Breast cancer outcome prediction with tumour tissue images and machine learning.

PURPOSE: Recent advances in machine learning have enabled better understanding of large and complex visual data. Here, we aim to investigate patient outcome prediction with a machine learning method using only an image of tumour sample as an input. METHODS: Utilising tissue microarray (TMA) samples obtained from the primary tumour of patients (N = 1299) within a nationwide breast cancer series with long-term-follow-up, we train and validate a machine learning method for patient outcome prediction. The prediction is performed by classifying samples into low or high digital risk score (DRS) groups. The outcome classifier is trained using sample images of 868 patients and evaluated and compared with human expert classification in a test set of 431 patients. RESULTS: In univariate survival analysis, the DRS classification resulted in a hazard ratio of 2.10 (95% CI 1.33-3.32, p = 0.001) for breast cancer-specific survival. The DRS classification remained as an independent predictor of breast cancer-specific survival in a multivariate Cox model with a hazard ratio of 2.04 (95% CI 1.20-3.44, p = 0.007). The accuracy (C-index) of the DRS grouping was 0.60 (95% CI 0.55-0.65), as compared to 0.58 (95% CI 0.53-0.63) for human expert predictions based on the same TMA samples. CONCLUSIONS: Our findings demonstrate the feasibility of learning prognostic signals in tumour tissue images without domain knowledge. Although further validation is needed, our study suggests that machine learning algorithms can extract prognostically relevant information from tumour histology complementing the currently used prognostic factors in breast cancer.

Using online game-based platforms to improve student performance and engagement in histology teaching.

BACKGROUND: Human morphology is a critical component of dental and medical graduate training. Innovations in basic science teaching methods are needed to keep up with an ever-changing landscape of technology. The purpose of this study was to investigate whether students in a medical and dental histology course would have better grades if they used gaming software Kahoot® and whether gamification effects on learning and enjoyment. METHODS: In an effort to both evoke students' interest and expand their skill retention, an online competition using Kahoot® was implemented for first-year students in 2018 (n = 215) at the University of Eastern Finland. Additionally, closed (160/215) or open-ended (41/215) feedback questions were collected and analyzed. RESULTS: The Kahoot® gamification program was successful and resulted in learning gains. The overall participant satisfaction using Kahoot® was high, with students (124/160) indicating that gamification increased their motivation to learn. The gaming approach seemed to enable the students to overcome individual difficulties (139/160) and to set up collaboration (107/160); furthermore, gamification promoted interest (109/160), and the respondents found the immediate feedback from senior professionals to be positive (146/160). In the open-ended survey, the students (23/41) viewed collaborative team- and gamification-based learning positively. CONCLUSION: This study lends support to the use of gamification in the teaching of histology and may provide a foundation for designing a gamification-integrated curriculum across healthcare disciplines.

UCHL5 expression associates with improved survival in lymph-node-positive rectal cancer.

Colorectal cancer is among the three most common cancer types for both genders, with a rising global incidence. To date, prognostic evaluation is difficult and largely dependent on early detection and successful surgery. UCHL5/Uch37 is an integral part of the protein homeostasis network as one of the three deubiquitinating enzymes associated with the 26S proteasome. Here, we have investigated in colorectal cancer the possible association of UCHL5 tumor expression and patient survival. UCHL5 tumor expression was evaluated by immunohistochemistry in 779 surgically treated colorectal cancer patients from Helsinki University Hospital, Finland, with assessment of clinicopathological parameters and the effect of UCHL5 expression on patient survival. High and undetectable UCHL5 expression both correlated with increased overall disease-specific survival in the subgroup of patients with lymph-node-positive (Dukes C/stage III) rectal cancer. Within this subgroup of 105 stage-III rectal cancer patients, none of the 7 with high UCHL5 expression died of colorectal cancer within 10 years after surgery ( p = 0.012). A similar, though less prominent, survival trend occurred throughout the whole patient cohort. In conclusion, UCHL5 is a promising novel prognostic marker in lymph-node-positive rectal cancer. Our results also advance the currently limited knowledge of biomarkers in colorectal cancer treatment.

Loss of PTEN expression in ERG-negative prostate cancer predicts secondary therapies and leads to shorter disease-specific survival time after radical prostatectomy.

The clinical course of prostate cancer is highly variable. Current prognostic variables, stage, and Gleason score have limitations in assessing treatment regimens for individual patients, especially in the intermediate-risk group of Gleason score 7. ERG:TMPRSS2 fusion and loss of PTEN are some of the most common genetic alterations in prostate cancer. Immunohistochemistry of PTEN and ERG has generated interest as a promising method for more precise outcome prediction but requires further validation in population-based cohorts. We studied the predictive value of ERG and PTEN expression by immunohistochemistry in two large radical prostatectomy cohorts comprising 815 patients with extensive follow-up information. Clinical end points were initiation of secondary therapy, overall survival, and disease-specific survival. Predictions of clinical outcomes were also assessed according to androgen receptor (AR) activity. PTEN loss, especially in ERG-negative cancers, predicted initiation of secondary treatments and shortened disease-specific survival time, as well as stratifying Gleason score 7 patients into different prognostic groups with regard to secondary treatments and disease-specific survival. High AR immunoreactivity in ERG-negative cancers with PTEN loss predicted worse disease-specific survival. We also observed that in Gleason score 7 ERG-negative cases with PTEN loss and high AR expression have significantly shorter disease-specific survival time compared with ERG-positive cases. Our conclusion is that loss of PTEN is a strong determining factor for shorter disease-specific survival time and initiation of secondary therapies after radical prostatectomy. The predictive value of PTEN immunoreactivity is further accentuated in ERG-negative cancers with high AR expression. Negative PTEN expression, accompanied by ERG status, can be used to stratify patients with Gleason score 7 into different survival groups. Assessment of PTEN and ERG status could provide an additional tool for initial diagnostics when determining the prognosis and subsequent follow-up regimen for patients treated by radical prostatectomy.

Androgen receptor-interacting protein HSPBAP1 facilitates growth of prostate cancer cells in androgen-deficient conditions.

Hormonal therapies targeting androgen receptor (AR) are effective in prostate cancer (PCa), but often the cancers progress to fatal castrate-resistant disease. Improved understanding of the cellular events during androgen deprivation would help to identify survival and stress pathways whose inhibition could synergize with androgen deprivation. Toward this aim, we performed an RNAi screen on 2,068 genes, including kinases, phosphatases, epigenetic enzymes and other druggable gene targets. High-content cell spot microarray (CSMA) screen was performed in VCaP cells in the presence and absence of androgens with detection of Ki67 and cleaved ADP-ribose polymerase (cPARP) as assays for cell proliferation and apoptosis. Thirty-nine candidate genes were identified, whose silencing inhibited proliferation or induced apoptosis of VCaP cells exclusively under androgen-deprived conditions. One of the candidates, HSPB (heat shock 27 kDa)-associated protein 1 (HSPBAP1), was confirmed to be highly expressed in tumor samples and its mRNA expression levels increased with the Gleason grade. We found that strong HSPBAP1 immunohistochemical staining (IHC) was associated with shorter disease-specific survival of PCa patients compared with negative to moderate staining. Furthermore, we demonstrate that HSPBAP1 interacts with AR in the nucleus of PCa cells specifically during androgen-deprived conditions, occupies chromatin at PSA/klk3 and TMPRSS2/tmprss2 enhancers and regulates their expression. In conclusion, we suggest that HSPBAP1 aids in sustaining cell viability by maintaining AR signaling during androgen-deprived conditions.

Dual role of FoxA1 in androgen receptor binding to chromatin, androgen signalling and prostate cancer.

High androgen receptor (AR) level in primary tumour predicts increased prostate cancer-specific mortality. However, the mechanisms that regulate AR function in prostate cancer are poorly known. We report here a new paradigm for the forkhead protein FoxA1 action in androgen signalling. Besides pioneering the AR pathway, FoxA1 depletion elicited extensive redistribution of AR-binding sites (ARBs) on LNCaP-1F5 cell chromatin that was commensurate with changes in androgen-dependent gene expression signature. We identified three distinct classes of ARBs and androgen-responsive genes: (i) independent of FoxA1, (ii) pioneered by FoxA1 and (iii) masked by FoxA1 and functional upon FoxA1 depletion. FoxA1 depletion also reprogrammed AR binding in VCaP cells, and glucocorticoid receptor binding and glucocorticoid-dependent signalling in LNCaP-1F5 cells. Importantly, FoxA1 protein level in primary prostate tumour had significant association to disease outcome; high FoxA1 level was associated with poor prognosis, whereas low FoxA1 level, even in the presence of high AR expression, predicted good prognosis. The role of FoxA1 in androgen signalling and prostate cancer is distinctly different from that in oestrogen signalling and breast cancer.

Diagnosis of soil-transmitted helminth infections with digital mobile microscopy and artificial intelligence in a resource-limited setting.

BACKGROUND: Infections caused by soil-transmitted helminths (STHs) are the most prevalent neglected tropical diseases and result in a major disease burden in low- and middle-income countries, especially in school-aged children. Improved diagnostic methods, especially for light intensity infections, are needed for efficient, control and elimination of STHs as a public health problem, as well as STH management. Image-based artificial intelligence (AI) has shown promise for STH detection in digitized stool samples. However, the diagnostic accuracy of AI-based analysis of entire microscope slides, so called whole-slide images (WSI), has previously not been evaluated on a sample-level in primary healthcare settings in STH endemic countries. METHODOLOGY/PRINCIPAL FINDINGS: Stool samples (n = 1,335) were collected during 2020 from children attending primary schools in Kwale County, Kenya, prepared according to the Kato-Katz method at a local primary healthcare laboratory and digitized with a portable whole-slide microscopy scanner and uploaded via mobile networks to a cloud environment. The digital samples of adequate quality (n = 1,180) were split into a training (n = 388) and test set (n = 792) and a deep-learning system (DLS) developed for detection of STHs. The DLS findings were compared with expert manual microscopy and additional visual assessment of the digital samples in slides with discordant results between the methods. Manual microscopy detected 15 (1.9%) Ascaris lumbricoides, 172 (21.7%) Tricuris trichiura and 140 (17.7%) hookworm (Ancylostoma duodenale or Necator americanus) infections in the test set. Importantly, more than 90% of all STH positive cases represented light intensity infections. With manual microscopy as the reference standard, the sensitivity of the DLS as the index test for detection of A. lumbricoides, T. trichiura and hookworm was 80%, 92% and 76%, respectively. The corresponding specificity was 98%, 90% and 95%. Notably, in 79 samples (10%) classified as negative by manual microscopy for a specific species, STH eggs were detected by the DLS and confirmed correct by visual inspection of the digital samples. CONCLUSIONS/SIGNIFICANCE: Analysis of digitally scanned stool samples with the DLS provided high diagnostic accuracy for detection of STHs. Importantly, a substantial number of light intensity infections were missed by manual microscopy but detected by the DLS. Thus, analysis of WSIs with image-based AI may provide a future tool for improved detection of STHs in a primary healthcare setting, which in turn could facilitate monitoring and evaluation of control programs.

External validation of a deep learning-based algorithm for detection of tall cells in papillary thyroid carcinoma: A multicenter study.

The tall cell subtype (TC-PTC) is an aggressive subtype of papillary thyroid carcinoma (PTC). The TC-PTC is defined as a PTC comprising at least 30% epithelial cells that are three times as tall as they are wide. In practice, this definition is difficult to adhere to, resulting in high inter-observer variability. In this multicenter study, we validated a previously trained deep learning (DL)-based algorithm for detection of tall cells on 160 externally collected hematoxylin and eosin (HE)-stained PTC whole-slide images. In a test set of 360 manual annotations of regions of interest from 18 separate tissue sections in the external dataset, the DL-based algorithm detected TCs with a sensitivity of 90.6% and a specificity of 88.5%. The DL algorithm detected non-TC areas with a sensitivity of 81.6% and a specificity of 92.9%. In the validation datasets, 20% and 30% TC thresholds correlated with a significantly shorter relapse-free survival. In conclusion, the DL algorithm detected TCs in unseen, external scanned HE tissue slides with high sensitivity and specificity without any retraining.

Development and evaluation of deep learning algorithms for assessment of acute burns and the need for surgery.

Assessment of burn extent and depth are critical and require very specialized diagnosis. Automated image-based algorithms could assist in performing wound detection and classification. We aimed to develop two deep-learning algorithms that respectively identify burns, and classify whether they require surgery. An additional aim assessed the performances in different Fitzpatrick skin types. Annotated burn (n = 1105) and background (n = 536) images were collected. Using a commercially available platform for deep learning algorithms, two models were trained and validated on 70% of the images and tested on the remaining 30%. Accuracy was measured for each image using the percentage of wound area correctly identified and F1 scores for the wound identifier; and area under the receiver operating characteristic (AUC) curve, sensitivity, and specificity for the wound classifier. The wound identifier algorithm detected an average of 87.2% of the wound areas accurately in the test set. For the wound classifier algorithm, the AUC was 0.885. The wound identifier algorithm was more accurate in patients with darker skin types; the wound classifier was more accurate in patients with lighter skin types. To conclude, image-based algorithms can support the assessment of acute burns with relatively good accuracy although larger and different datasets are needed.

An extensive tumor array analysis supports tumor suppressive role for nucleophosmin in breast cancer.

Nucleophosmin (NPM) is a multifunctional protein involved in a complex network of interactions. The role of NPM in oncogenesis is controversial. The NPM gene (NPM1) is mutated or rearranged in a number of hematological disorders, but such changes have not been detected in solid cancers. However, experiments with cultured NPM-null cells and with mice carrying a single inactivated NPM allele indicate a tumor suppressor function for NPM. To resolve the role of NPM in solid cancers, we examined its expression and localization in histologically normal breast tissue and a large array of human breast carcinoma samples (n = 1160), and also evaluated its association with clinicopathological variables and patient survival. The intensity and localization (nucleolar, nuclear, cytoplasmic) of NPM varied across clinical samples. No mutations explaining the differences were found, but the present findings indicate that expression levels of NPM affected its localization. Our study also revealed a novel granular staining pattern for NPM, which was an independent prognostic factor of poor prognosis. In addition, reduced levels of NPM protein were associated with poor prognosis. Furthermore, luminal epithelial cells of histologically normal breast displayed high levels of NPM and overexpression of NPM in the invasive MDA-MB-231 cells abrogated their growth in soft agar. These results support a tumor suppressive role for NPM in breast cancer.

Decreased xanthine oxidoreductase (XOR) is associated with a worse prognosis in patients with serous ovarian carcinoma.

OBJECTIVE: Xanthine oxidoreductase (XOR) is a key enzyme in the degradation of DNA, RNA and high-energy phosphates. In the human cancers previously studied, down-regulated XOR identifies patients with unfavorable prognosis. We assessed the clinical relevance of XOR expression in serous ovarian cancer. METHODS: XOR protein was determined in tissue microarrays from 474 patients with serous ovarian cancer and analyzed with respect to clinical parameters and survival. RESULTS: XOR was down regulated in 64% of the tumors as compared to the corresponding normal tissue. Decreased XOR was associated with a poorly differentiated tumor and an abnormal p53 expression, but not with age at diagnosis, FIGO stage, Ki-67 or tumor size. XOR expression was associated with outcome, and the five year ovarian cancer specific survival in patients with strong XOR expression was 59% compared to 44% in those with moderate (hazard ratio, HR; 1.44; P=0.0083) and 26% in patients with lack of XOR (HR, 2.07; P=0.0003). This was also true in patients whose tumors were highly differentiated (HR, 3.67; P=0.008) and in patients with a small (<1cm) residual tumor (HR, 2.62; P=0.017), and in patients whose tumors show a low Ki-67 protein expression (HR, 3.79; P<0.0001). In multivariate survival analysis, absence of XOR emerged as an independent prognostic factor (HR, 1.82; P=0.015). CONCLUSIONS: Decreased XOR is associated with poorer prognosis in patients with serous ovarian cancer especially in those with an otherwise more favorable prognostic profile.

Lack of MMP-9 expression is a marker for poor prognosis in Dukes' B colorectal cancer.

BACKGROUND: Matrix metalloproteinases (MMPs) play a role in cancer progression by degrading extracellular matrix and basement membranes, assisting in tumour neovascularization and in supporting immune response in cancer. METHODS: We studied the prognostic value of immunohistochemical expression of MMP-2, MMP-8, and MMP-9 in a series of 619 colorectal cancer patients using tissue microarray specimens. RESULTS: Of the samples, 56% were positive for MMP-2, 78% for MMP-8, and 60% for MMP-9. MMP-9 associated with low WHO grade (p < 0.001). In univariate analysis of Dukes' B tumours, MMP-9 negativity associated with poor survival (p = 0.018), and MMP-9 positivity was an independent prognostic marker in multivariate analysis of these tumours (p = 0.034). CONCLUSION: Negative MMP-9 expression can predict poor prognosis in Dukes' B colorectal tumours and may prove useful for identifying patients, who should be offered adjuvant treatment.

A deep learning-based algorithm for tall cell detection in papillary thyroid carcinoma.

INTRODUCTION: According to the World Health Organization, the tall cell variant (TCV) is an aggressive subtype of papillary thyroid carcinoma (PTC) comprising at least 30% epithelial cells two to three times as tall as they are wide. In practice, applying this definition is difficult causing substantial interobserver variability. We aimed to train a deep learning algorithm to detect and quantify the proportion of tall cells (TCs) in PTC. METHODS: We trained the deep learning algorithm using supervised learning, testing it on an independent dataset, and further validating it on an independent set of 90 PTC samples from patients treated at the Hospital District of Helsinki and Uusimaa between 2003 and 2013. We compared the algorithm-based TC percentage to the independent scoring by a human investigator and how those scorings associated with disease outcomes. Additionally, we assessed the TC score in 71 local and distant tumor relapse samples from patients with aggressive disease. RESULTS: In the test set, the deep learning algorithm detected TCs with a sensitivity of 93.7% and a specificity of 94.5%, whereas the sensitivity fell to 90.9% and specificity to 94.1% for non-TC areas. In the validation set, the deep learning algorithm TC scores correlated with a diminished relapse-free survival using cutoff points of 10% (p = 0.044), 20% (p < 0.01), and 30% (p = 0.036). The visually assessed TC score did not statistically significantly predict survival at any of the analyzed cutoff points. We observed no statistically significant difference in the TC score between primary tumors and relapse tumors determined by the deep learning algorithm or visually. CONCLUSIONS: We present a novel deep learning-based algorithm to detect tall cells, showing that a high deep learning-based TC score represents a statistically significant predictor of less favorable relapse-free survival in PTC.

Outcome and Biomarker Supervised Deep Learning for Survival Prediction in Two Multicenter Breast Cancer Series.

BACKGROUND: Prediction of clinical outcomes for individual cancer patients is an important step in the disease diagnosis and subsequently guides the treatment and patient counseling. In this work, we develop and evaluate a joint outcome and biomarker supervised (estrogen receptor expression and ERBB2 expression and gene amplification) multitask deep learning model for prediction of outcome in breast cancer patients in two nation-wide multicenter studies in Finland (the FinProg and FinHer studies). Our approach combines deep learning with expert knowledge to provide more accurate, robust, and integrated prediction of breast cancer outcomes. MATERIALS AND METHODS: Using deep learning, we trained convolutional neural networks (CNNs) with digitized tissue microarray (TMA) samples of primary hematoxylin-eosin-stained breast cancer specimens from 693 patients in the FinProg series as input and breast cancer-specific survival as the endpoint. The trained algorithms were tested on 354 TMA patient samples in the same series. An independent set of whole-slide (WS) tumor samples from 674 patients in another multicenter study (FinHer) was used to validate and verify the generalization of the outcome prediction based on CNN models by Cox survival regression and concordance index (c-index). Visual cancer tissue characterization, i.e., number of mitoses, tubules, nuclear pleomorphism, tumor-infiltrating lymphocytes, and necrosis was performed on TMA samples in the FinProg test set by a pathologist and combined with deep learning-based outcome prediction in a multitask algorithm. RESULTS: The multitask algorithm achieved a hazard ratio (HR) of 2.0 (95% confidence interval [CI] 1.30-3.00), P < 0.001, c-index of 0.59 on the 354 test set of FinProg patients, and an HR of 1.7 (95% CI 1.2-2.6), P = 0.003, c-index 0.57 on the WS tumor samples from 674 patients in the independent FinHer series. The multitask CNN remained a statistically independent predictor of survival in both test sets when adjusted for histological grade, tumor size, and axillary lymph node status in a multivariate Cox analyses. An improved accuracy (c-index 0.66) was achieved when deep learning was combined with the tissue characteristics assessed visually by a pathologist. CONCLUSIONS: A multitask deep learning algorithm supervised by both patient outcome and biomarker status learned features in basic tissue morphology predictive of survival in a nationwide, multicenter series of patients with breast cancer. The algorithms generalized to another independent multicenter patient series and whole-slide breast cancer samples and provide prognostic information complementary to that of a comprehensive series of established prognostic factors.

Long-term prognosis of breast cancer detected by mammography screening or other methods.

INTRODUCTION: Previous studies of breast cancer have shown that patients whose tumors are detected by mammography screening have a more favorable survival. Little is known, however, about the long-term prognostic impact of screen detection. The purpose of the current study was to compare breast cancer-specific long-term survival of patients whose tumors were detected in mammography screening compared with those whose tumors were detected by other methods. METHODS: Breast cancer patients diagnosed within five specified geographical areas in Finland in 1991 and 1992 were identified (N = 2,936). Detailed clinical, treatment and outcome data, as well as tissue samples, were collected. Women with in situ carcinoma, distant metastases at the time of primary diagnosis and women who were not treated surgically were excluded. The main analyses were performed after excluding patients with other malignancy or contralateral breast cancer, followed by sensitivity analyses with different exclusion criteria. Median follow-up time was 15.4 years. Univariate and multivariate analyses of breast cancer-specific survival were performed. RESULTS: Of patients included in the main analyses (n = 1,884), 22% (n = 408) of cancers were screen-detected and 78% (n = 1,476) were detected by other methods. Breast cancer-specific 15-year survival was 86% for patients with screen-detected cancer and 66% for patients diagnosed using other methods (P < 0.0001, HR = 2.91). Similar differences in survival were observed in women at screening age (50 to 69 years), as well as in clinically important subgroups, such as patients with small tumors (≤ 1 cm in diameter) and without nodal involvement (N0). Women with breast cancer diagnosed on the basis of screening mammography had a more favorable prognosis than those diagnosed outside screening programs, following adjustments according to patient age, tumor size, axillary lymph node status, histological grade and hormone receptor status. Significant differences in the risk of having future contralateral breast cancer according to method of detection were not observed. CONCLUSIONS: Breast cancer detected by mammography screening is an independent prognostic factor in breast cancer and is associated with a more favorable survival rate as well as in long-term follow-up.

Breast cancer biological subtypes and protein expression predict for the preferential distant metastasis sites: a nationwide cohort study.

INTRODUCTION: Some molecular subtypes of breast cancer have preferential sites of distant relapse. The protein expression pattern of the primary tumor may influence the first distant metastasis site. METHODS: We identified from the files of the Finnish Cancer Registry patients diagnosed with breast cancer in five geographical regions Finland in 1991-1992, reviewed the hospital case records, and collected primary tumor tissue. Out of the 2,032 cases identified, 234 developed distant metastases after a median follow-up time of 2.7 years and had the first metastatic site documented (a total of 321 sites). Primary tumor microarray (TMA) cores were analyzed for 17 proteins using immunohistochemistry and for erbB2 using chromogenic in situ hybridization, and their associations with the first metastasis site were examined. The cancers were classified into luminal A, luminal B, HER2+ enriched, basal-like or non-expressor subtypes. RESULTS: A total of 3,886 TMA cores were analyzed. Luminal A cancers had a propensity to give rise first to bone metastases, HER2-enriched cancers to liver and lung metastases, and basal type cancers to liver and brain metastases. Primary tumors that gave first rise to bone metastases expressed frequently estrogen receptor (ER) and SNAI1 (SNAIL) and rarely COX2 and HER2, tumors with first metastases in the liver expressed infrequently SNAI1, those with lung metastases expressed frequently the epidermal growth factor receptor (EGFR), cytokeratin-5 (CK5) and HER2, and infrequently progesterone receptor (PgR), tumors with early skin metastases expressed infrequently E-cadherin, and breast tumors with first metastases in the brain expressed nestin, prominin-1 and CK5 and infrequently ER and PgR. CONCLUSIONS: Breast tumor biological subtypes have a tendency to give rise to first distant metastases at certain body sites. Several primary tumor proteins were associated with homing of breast cancer cells.

Development and evaluation of a virtual microscopy application for automated assessment of Ki-67 expression in breast cancer.

BACKGROUND: The aim of the study was to develop a virtual microscopy enabled method for assessment of Ki-67 expression and to study the prognostic value of the automated analysis in a comprehensive series of patients with breast cancer. METHODS: Using a previously reported virtual microscopy platform and an open source image processing tool, ImageJ, a method for assessment of immunohistochemically (IHC) stained area and intensity was created. A tissue microarray (TMA) series of breast cancer specimens from 1931 patients was immunostained for Ki-67, digitized with a whole slide scanner and uploaded to an image web server. The extent of Ki-67 staining in the tumour specimens was assessed both visually and with the image analysis algorithm. The prognostic value of the computer vision assessment of Ki-67 was evaluated by comparison of distant disease-free survival in patients with low, moderate or high expression of the protein. RESULTS: 1648 evaluable image files from 1334 patients were analysed in less than two hours. Visual and automated Ki-67 extent of staining assessments showed a percentage agreement of 87% and weighted kappa value of 0.57. The hazard ratio for distant recurrence for patients with a computer determined moderate Ki-67 extent of staining was 1.77 (95% CI 1.31-2.37) and for high extent 2.34 (95% CI 1.76-3.10), compared to patients with a low extent. In multivariate survival analyses, automated assessment of Ki-67 extent of staining was retained as a significant prognostic factor. CONCLUSIONS: Running high-throughput automated IHC algorithms on a virtual microscopy platform is feasible. Comparison of visual and automated assessments of Ki-67 expression shows moderate agreement. In multivariate survival analysis, the automated assessment of Ki-67 extent of staining is a significant and independent predictor of outcome in breast cancer.

Deep learning identifies morphological features in breast cancer predictive of cancer ERBB2 status and trastuzumab treatment efficacy.

The treatment of patients with ERBB2 (HER2)-positive breast cancer with anti-ERBB2 therapy is based on the detection of ERBB2 gene amplification or protein overexpression. Machine learning (ML) algorithms can predict the amplification of ERBB2 based on tumor morphological features, but it is not known whether ML-derived features can predict survival and efficacy of anti-ERBB2 treatment. In this study, we trained a deep learning model with digital images of hematoxylin-eosin (H&E)-stained formalin-fixed primary breast tumor tissue sections, weakly supervised by ERBB2 gene amplification status. The gene amplification was determined by chromogenic in situ hybridization (CISH). The training data comprised digitized tissue microarray (TMA) samples from 1,047 patients. The correlation between the deep learning-predicted ERBB2 status, which we call H&E-ERBB2 score, and distant disease-free survival (DDFS) was investigated on a fully independent test set, which included whole-slide tumor images from 712 patients with trastuzumab treatment status available. The area under the receiver operating characteristic curve (AUC) in predicting gene amplification in the test sets was 0.70 (95% CI, 0.63-0.77) on 354 TMA samples and 0.67 (95% CI, 0.62-0.71) on 712 whole-slide images. Among patients with ERBB2-positive cancer treated with trastuzumab, those with a higher than the median morphology-based H&E-ERBB2 score derived from machine learning had more favorable DDFS than those with a lower score (hazard ratio [HR] 0.37; 95% CI, 0.15-0.93; P = 0.034). A high H&E-ERBB2 score was associated with unfavorable survival in patients with ERBB2-negative cancer as determined by CISH. ERBB2-associated morphology correlated with the efficacy of adjuvant anti-ERBB2 treatment and can contribute to treatment-predictive information in breast cancer.

Point-of-Care Digital Cytology With Artificial Intelligence for Cervical Cancer Screening in a Resource-Limited Setting.

Importance: Cervical cancer is highly preventable but remains a common and deadly cancer in areas without screening programs. The creation of a diagnostic system to digitize Papanicolaou test samples and analyze them using a cloud-based deep learning system (DLS) may provide needed cervical cancer screening to resource-limited areas. Objective: To determine whether artificial intelligence-supported digital microscopy diagnostics can be implemented in a resource-limited setting and used for analysis of Papanicolaou tests. Design, Setting, and Participants: In this diagnostic study, cervical smears from 740 HIV-positive women aged between 18 and 64 years were collected between September 1, 2018, and September 30, 2019. The smears were digitized with a portable slide scanner, uploaded to a cloud server using mobile networks, and used to train and validate a DLS for the detection of atypical cervical cells. This single-center study was conducted at a local health care center in rural Kenya. Exposures: Detection of squamous cell atypia in the digital samples by analysis with the DLS. Main Outcomes and Measures: The accuracy of the DLS in the detection of low- and high-grade squamous intraepithelial lesions in Papanicolaou test whole-slide images. Results: Papanicolaou test results from 740 HIV-positive women (mean [SD] age, 41.8 [10.3] years) were collected. The DLS was trained using 350 whole-slide images and validated on 361 whole-slide images (average size, 100 387 × 47 560 pixels). For detection of cervical cellular atypia, sensitivities were 95.7% (95% CI, 85.5%-99.5%) and 100% (95% CI, 82.4%-100%), and specificities were 84.7% (95% CI, 80.2%-88.5%) and 78.4% (95% CI, 73.6%-82.4%), compared with the pathologist assessment of digital and physical slides, respectively. Areas under the receiver operating characteristic curve were 0.94 and 0.96, respectively. Negative predictive values were high (99%-100%), and accuracy was high, particularly for the detection of high-grade lesions. Interrater agreement was substantial compared with the pathologist assessment of digital slides (κ = 0.72) and fair compared with the assessment of glass slides (κ = 0.36). No samples that were classified as high grade by manual sample analysis had false-negative assessments by the DLS. Conclusions and Relevance: In this study, digital microscopy with artificial intelligence was implemented at a rural clinic and used to detect atypical cervical smears with a high sensitivity compared with visual sample analysis.