Encrypted federated learning for secure decentralized collaboration in cancer image analysis.

Artificial intelligence (AI) has a multitude of applications in cancer research and oncology. However, the training of AI systems is impeded by the limited availability of large datasets due to data protection requirements and other regulatory obstacles. Federated and swarm learning represent possible solutions to this problem by collaboratively training AI models while avoiding data transfer. However, in these decentralized methods, weight updates are still transferred to the aggregation server for merging the models. This leaves the possibility for a breach of data privacy, for example by model inversion or membership inference attacks by untrusted servers. Somewhat-homomorphically-encrypted federated learning (SHEFL) is a solution to this problem because only encrypted weights are transferred, and model updates are performed in the encrypted space. Here, we demonstrate the first successful implementation of SHEFL in a range of clinically relevant tasks in cancer image analysis on multicentric datasets in radiology and histopathology. We show that SHEFL enables the training of AI models which outperform locally trained models and perform on par with models which are centrally trained. In the future, SHEFL can enable multiple institutions to co-train AI models without forsaking data governance and without ever transmitting any decryptable data to untrusted servers.

Transformer-based biomarker prediction from colorectal cancer histology: A large-scale multicentric study.

Deep learning (DL) can accelerate the prediction of prognostic biomarkers from routine pathology slides in colorectal cancer (CRC). However, current approaches rely on convolutional neural networks (CNNs) and have mostly been validated on small patient cohorts. Here, we develop a new transformer-based pipeline for end-to-end biomarker prediction from pathology slides by combining a pre-trained transformer encoder with a transformer network for patch aggregation. Our transformer-based approach substantially improves the performance, generalizability, data efficiency, and interpretability as compared with current state-of-the-art algorithms. After training and evaluating on a large multicenter cohort of over 13,000 patients from 16 colorectal cancer cohorts, we achieve a sensitivity of 0.99 with a negative predictive value of over 0.99 for prediction of microsatellite instability (MSI) on surgical resection specimens. We demonstrate that resection specimen-only training reaches clinical-grade performance on endoscopic biopsy tissue, solving a long-standing diagnostic problem.

Generalizable biomarker prediction from cancer pathology slides with self-supervised deep learning: A retrospective multi-centric study.

Deep learning (DL) can predict microsatellite instability (MSI) from routine histopathology slides of colorectal cancer (CRC). However, it is unclear whether DL can also predict other biomarkers with high performance and whether DL predictions generalize to external patient populations. Here, we acquire CRC tissue samples from two large multi-centric studies. We systematically compare six different state-of-the-art DL architectures to predict biomarkers from pathology slides, including MSI and mutations in BRAF, KRAS, NRAS, and PIK3CA. Using a large external validation cohort to provide a realistic evaluation setting, we show that models using self-supervised, attention-based multiple-instance learning consistently outperform previous approaches while offering explainable visualizations of the indicative regions and morphologies. While the prediction of MSI and BRAF mutations reaches a clinical-grade performance, mutation prediction of PIK3CA, KRAS, and NRAS was clinically insufficient.

Swarm learning for decentralized artificial intelligence in cancer histopathology.

Artificial intelligence (AI) can predict the presence of molecular alterations directly from routine histopathology slides. However, training robust AI systems requires large datasets for which data collection faces practical, ethical and legal obstacles. These obstacles could be overcome with swarm learning (SL), in which partners jointly train AI models while avoiding data transfer and monopolistic data governance. Here, we demonstrate the successful use of SL in large, multicentric datasets of gigapixel histopathology images from over 5,000 patients. We show that AI models trained using SL can predict BRAF mutational status and microsatellite instability directly from hematoxylin and eosin (H&E)-stained pathology slides of colorectal cancer. We trained AI models on three patient cohorts from Northern Ireland, Germany and the United States, and validated the prediction performance in two independent datasets from the United Kingdom. Our data show that SL-trained AI models outperform most locally trained models, and perform on par with models that are trained on the merged datasets. In addition, we show that SL-based AI models are data efficient. In the future, SL can be used to train distributed AI models for any histopathology image analysis task, eliminating the need for data transfer.

Prognostic and Predictive Value of Tumor Budding in Colorectal Cancer.

BACKGROUND: Tumor budding (TB) is an adverse prognostic factor in colorectal cancer (CRC). International consensus on a standardized assessment method has led to its wider reporting. However, uncertainty regarding its clinical value persists. This study aimed to (1) confirm the prognostic significance of TB, particularly in stage II CRC; (2) to determine optimum thresholds for TB risk grouping; and (3) to determine whether TB influences responsiveness to chemotherapy. METHODS: TB was assessed in CRC sections from 1575 QUASAR trial patients randomized between adjuvant chemotherapy and observation. Optimal risk group cutoffs were determined by maximum likelihood methods, with their influence on recurrence and mortality investigated in stratified log-rank analyses on exploratory (n = 504), hypothesis-testing (n = 478), and final (n = 593) data sets. RESULTS: The optimal threshold for high-grade TB (HGTB) was ≥ 10 buds per 1.23 mm2. High-grade TB tumors had significantly worse outcomes than those with lower TB: 10-year recurrence 36% versus 22% (risk ratio, 2.00 [95% CI, 1.62-2.45]; 2P < .0001) and 10-year mortality 50% vs. 37% (risk ratio, 1.53 [95% CI, 1.34-1.76]; 2P < .0001). The prognostic significance remained equally strong after allowance for other pathological risk factors, including stage, grade, lymphovascular invasion, and mismatch repair status. There was a nonsignificant trend toward increasing chemotherapy efficacy with increasing bud counts. CONCLUSIONS: TB is a strong independent predictor of recurrence. Chemotherapy efficacy is comparable in patients with higher and lower TB; hence, absolute reductions in recurrence and death with chemotherapy should be about twice as large in patients with ≥ 10 than < 10 TB counts.

Lynch syndrome screening in colorectal cancer: results of a prospective 2-year regional programme validating the NICE diagnostics guidance pathway throughout a 5.2-million population.

AIMS: Screening all patients newly diagnosed with colorectal cancer (CRC) for possible Lynch syndrome (LS) has been recommended in the United Kingdom since the National Institute for Health and Care Excellence (NICE) released new diagnostics guidance in February 2017. We sought to validate the NICE screening pathway through a prospective regional programme throughout a 5.2-million population during a 2-year period. METHODS AND RESULTS: Pathology departments at 14 hospital trusts in the Yorkshire and Humber region of the United Kingdom were invited to refer material from patients with newly diagnosed CRC aged 50 years or over between 1 April 2017 and 31 March 2019 for LS screening. Testing consisted of immunohistochemistry for MLH1, PMS2, MSH2 and MSH6 followed by BRAF mutation analysis ± MLH1 promoter methylation testing in cases showing MLH1 loss. A total of 3141 individual specimens were submitted for testing from 12 departments consisting of 3061 unique tumours and 2791 prospectively acquired patients with CRC. Defective mismatch repair (dMMR) was observed in 15% of cases. In cases showing MLH1 loss, 76% contained a detectable BRAF mutation and, of the remainder, 77% showed MLH1 promoter hypermethylation. Of the patients included in the final analysis, 81 (2.9%) had an indication for germline testing. CONCLUSION: LS screening using the NICE diagnostics guidance pathway is deliverable at scale identifying significant numbers of patients with dMMR. This information is used to refer patients to regional clinical genetics services in addition to informing treatment pathways including the use of adjuvant/neoadjuvant chemotherapy and immunotherapy.

Development of a Remote Online Collaborative Medical School Pathology Curriculum with Clinical Correlations, across Several International Sites, through the Covid-19 Pandemic.

INTRODUCTION: Due to the Covid-19 social distancing restrictions, in March 2020, Weill Cornell Medicine-Qatar decided to replace students' clinical instruction with novel online electives. Hence, we implemented an innovative online and remote pathology curriculum, anchored on virtual microscopy and Zoom videoconferencing: ideal tools to support online teaching. OBJECTIVE: To assess a new curriculum implementation at Weill Cornell Medicine-Qatar. MATERIALS AND METHODS: This for-credit, 2-week elective included 6 synchronous Zoom sessions where complex clinicopathological cases were discussed in small groups. We used open access digital microscopy slides from the University of Leeds' Virtual Pathology Library (http://www.virtualpathology.leeds.ac.uk/slides/library/). Students independently prepared for these sessions by reviewing cases, slides, readings, and questions in advance (asynchronous self-directed learning anchored on a flipped classroom model), and wrote a final review of a case. An assessment and feedback were given to each student. RESULTS: Four elective iterations were offered to a total of 29 students, with learners and faculty spread over 4 countries. During the Zoom sessions, students controlled the digital slides and offered their own diagnoses, followed by group discussions to strengthen autonomy and confidence. We surveyed learners about the elective's performance (program evaluation). Students conveyed high levels of satisfaction about the elective's overall quality, their pathology learning and online interactions, with minimal challenges related to the remote nature of the course. DISCUSSION AND CONCLUSIONS: Technological innovations mitigate sudden disruptions in medical education. A remote curriculum allows instruction at any distance, at any time, from anywhere, enhancing educational exchanges, flexibility and globalization in medical education.

The Effect of Quality Control on Accuracy of Digital Pathology Image Analysis.

Digital slide images produced from routine diagnostic histopathological preparations suffer from variation arising at every step of the processing pipeline. Typically, pathologists compensate for such variation using expert knowledge and experience, which is difficult to replicate in automated solutions. The extent to which inconsistencies affect image analysis is explored in this work, examining in detail, the results from a previously published algorithm automating the generation of tumor:stroma ratio (TSR) in colorectal clinical trial datasets. One dataset consisting of 2,211 cases and 106,268 expert-labelled images is used to identify quality issues, by visually inspecting cases where algorithm-pathologist agreement is lowest. Twelve categories are identified and used to analyze pathologist-algorithm agreement in relation to these categories. Of the 2,211 cases, 701 were found to be free from any image quality issues. Algorithm performance was then assessed, comparing pathologist agreement with image quality classification. It was found that agreement was lowest on poorly differentiated tissue, with a mean TSR difference of 0.25 (sd = 0.24). Removing images that contained quality issues increased accuracy from 80% to 83%, at the expense of reducing the dataset to 33,736 images (32%). Training the algorithm on the optimized dataset, prior to testing on all images saw a decrease in accuracy of 4%, indicating that the optimized dataset did not contain enough variation to generate a fully representative model. The results provide an in-depth perspective on image quality, highlighting the importance of the effects on downstream image analysis.

Clinical-Grade Detection of Microsatellite Instability in Colorectal Tumors by Deep Learning.

BACKGROUND & AIMS: Microsatellite instability (MSI) and mismatch-repair deficiency (dMMR) in colorectal tumors are used to select treatment for patients. Deep learning can detect MSI and dMMR in tumor samples on routine histology slides faster and less expensively than molecular assays. However, clinical application of this technology requires high performance and multisite validation, which have not yet been performed. METHODS: We collected H&E-stained slides and findings from molecular analyses for MSI and dMMR from 8836 colorectal tumors (of all stages) included in the MSIDETECT consortium study, from Germany, the Netherlands, the United Kingdom, and the United States. Specimens with dMMR were identified by immunohistochemistry analyses of tissue microarrays for loss of MLH1, MSH2, MSH6, and/or PMS2. Specimens with MSI were identified by genetic analyses. We trained a deep-learning detector to identify samples with MSI from these slides; performance was assessed by cross-validation (N = 6406 specimens) and validated in an external cohort (n = 771 specimens). Prespecified endpoints were area under the receiver operating characteristic (AUROC) curve and area under the precision-recall curve (AUPRC). RESULTS: The deep-learning detector identified specimens with dMMR or MSI with a mean AUROC curve of 0.92 (lower bound, 0.91; upper bound, 0.93) and an AUPRC of 0.63 (range, 0.59-0.65), or 67% specificity and 95% sensitivity, in the cross-validation development cohort. In the validation cohort, the classifier identified samples with dMMR with an AUROC of 0.95 (range, 0.92-0.96) without image preprocessing and an AUROC of 0.96 (range, 0.93-0.98) after color normalization. CONCLUSIONS: We developed a deep-learning system that detects colorectal cancer specimens with dMMR or MSI using H&E-stained slides; it detected tissues with dMMR with an AUROC of 0.96 in a large, international validation cohort. This system might be used for high-throughput, low-cost evaluation of colorectal tissue specimens.

Histological intratumoral heterogeneity in pretreatment esophageal cancer biopsies predicts survival benefit from neoadjuvant chemotherapy: results from the UK MRC OE02 trial.

Despite the use of multimodal treatment, survival of esophageal cancer (EC) patients remains poor. One proposed explanation for the relatively poor response to cytotoxic chemotherapy is intratumor heterogeneity. The aim was to establish a statistical model to objectively measure intratumor heterogeneity of the proportion of tumor (IHPoT) and to use this newly developed method to measure IHPoT in the pretreatment biopsies from from EC patients recruited to the OE02 trial. A statistical mixed effect model (MEM) was established for estimating IHPoT based on variation in hematoxylin/eosin (HE) stained pretreatment biopsy pieces from the same individual in 218 OE02 trial patients (103 treated by chemotherapy and surgery (chemo+surgery); 115 patients treated by surgery alone). The relationship between IHPoT, prognosis, chemotherapy survival benefit, and clinicopathological variables was assessed. About 97 (44.5%) and 121 (55.5%) ECs showed high and low IHPoT, respectively. There was no significant difference in IHPoT between surgery (median [range], 0.1637 [0-3.17]) and chemo+surgery (median [range], 0.1692 [0-2.69]) patients (P = 0.43). Chemo+surgery patients with low IHPoT had a significantly longer survival than surgery patients (HR = 1.81, 95% CI: 1.20-2.75, P = 0.005). There was no survival difference between chemo+surgery and surgery patients with high IHPoT (HR = 1.15, 95% CI: 0.72-1.81, P = 0.566). This is the first study suggesting that IHPoT measured in the pretreatment biopsy can predict chemotherapy survival benefit in EC patients. IHPoT may represent a clinically useful biomarker for patient treatment stratification. Future studies should determine if pathologists can reliably estimate IHPoT.

Confirmation that somatic mutations of beta-2 microglobulin correlate with a lack of recurrence in a subset of stage II mismatch repair deficient colorectal cancers from the QUASAR trial.

AIMS: Beta2-microglobulin (B2M) forms part of the HLA class I complex and plays a role in metastatic biology. B2M mutations occur frequently in mismatch repair-deficient colorectal cancer (dMMR CRC), with limited data suggesting they may protect against recurrence. Our experimental study tested this hypothesis by investigating B2M mutation status and B2M protein expression and recurrence in patients in the stage II QUASAR clinical trial. METHODS AND RESULTS: Sanger sequencing was performed for the three coding exons of B2M on 121 dMMR and a subsample of 108 pMMR tumours; 52 with recurrence and 56 without. B2M protein expression was assessed by immunohistochemistry. Mutation status and protein expression were correlated with recurrence and compared to proficient mismatch repair (pMMR) CRCs. Deleterious B2M mutations were detected in 39 of 121 (32%) dMMR tumours. Five contained missense B2M-variants of unknown significance, so were excluded from further analyses. With median follow-up of 7.4 years, none of the 39 B2M-mutant tumours recurred, compared with 14 of 77 (18%) B2M-wild-type tumours (P = 0.005); six at local and eight at distant sites. Sensitivity and specificity of IHC in detecting B2M mutations was 87 and 71%, respectively. Significantly (P < 0.0001) fewer (three of 104, 2.9%) of the 108 pMMR CRCs demonstrated deleterious B2M mutations. One pMMR tumour, containing a frameshift mutation, later recurred. CONCLUSION: B2M mutations were detected in nearly one-third of dMMR cancers, none of which recurred. B2M mutation status has potential clinical utility as a prognostic biomarker in stage II dMMR CRC. The mechanism of protection against recurrence and whether this protection extends to stage III disease remains unclear.

Prognostic value of pathological lymph node status and primary tumour regression grading following neoadjuvant chemotherapy - results from the MRC OE02 oesophageal cancer trial.

AIMS: Neoadjuvant chemotherapy (NAC) remains an important therapeutic option for advanced oesophageal cancer (OC). Pathological tumour regression grade (TRG) may offer additional information by directing adjuvant treatment and/or follow-up but its clinical value remains unclear. We analysed the prognostic value of TRG and associated pathological factors in OC patients enrolled in the Medical Research Council (MRC) OE02 trial. METHODS AND RESULTS: Histopathology was reviewed in 497 resections from OE02 trial participants randomised to surgery (S group; n = 244) or NAC followed by surgery [chemotherapy plus surgery (CS) group; n = 253]. The association between TRG groups [responders (TRG1-3) versus non-responders (TRG4-5)], pathological lymph node (LN) status and overall survival (OS) was analysed. One hundred and ninety-five of 253 (77%) CS patients were classified as 'non-responders', with a significantly higher mortality risk compared to responders [hazard ratio (HR) = 1.53, 95% confidence interval (CI) = 1.05-2.24, P = 0.026]. OS was significantly better in patients without LN metastases irrespective of TRG [non-responders HR = 1.87, 95% CI = 1.33-2.63, P < 0.001 versus responders HR = 2.21, 95% CI = 1.11-4.10, P = 0.024]. In multivariate analyses, LN status was the only independent factor predictive of OS in CS patients (HR = 1.93, 95% CI = 1.42-2.62, P < 0.001). Exploratory subgroup analyses excluding radiotherapy-exposed patients (n = 48) showed similar prognostic outcomes. CONCLUSION: Lymph node status post-NAC is the most important prognostic factor in patients with resectable oesophageal cancer, irrespective of TRG. Potential clinical implications, e.g. adjuvant treatment or intensified follow-up, reinforce the importance of LN dissection for staging and prognostication.

Intratumoral stromal morphometry predicts disease recurrence but not response to 5-fluorouracil-results from the QUASAR trial of colorectal cancer.

AIMS: The biological importance of tumour-associated stroma is becoming increasingly apparent, but its clinical utility remains ill-defined. For stage II/Dukes B colorectal cancer (CRC), clinical biomarkers are urgently required to direct therapeutic options. We report here prognostic/predictive analyses, and molecular associations, of stromal morphometric quantification in the Quick and Simple and Reliable (QUASAR) trial of CRC. METHODS AND RESULTS: Relative proportions of tumour epithelium (PoT) or stroma (PoS) were morphometrically quantified on digitised haematoxylin and eosin (H&E) sections derived from 1800 patients enrolled in QUASAR, which randomised 3239 (91% stage II) CRC patients between adjuvant fluorouracil/folinic acid (FUFA) chemotherapy and observation. The prognostic and predictive values of PoT/PoS measurements were determined by the use of stratified log-rank analyses. A high proportion of tumour stroma (≥50%) was associated with an increased recurrence risk: 31.3% (143/457) recurrence for ≥50% versus 21.9% (294/1343) for <50% [rate ratio (RR) 1.62; 95% confidence interval (CI) 1.30-2.02; P < 0.0001]. Of patients with stromal proportions of ≥65%, 40% (46/115) had recurrent disease within 10 years. The adverse prognostic effect of a high stromal proportion was independent of established prognostic variables, and was maintained in stage II/Dukes B patients (RR 1.62; 95% CI 1.26-2.08; P = 0.0002). KRAS mutation in the presence of a high stromal proportion augmented recurrence risk (RR 2.93; 95% CI 1.87-4.59; P = 0.0005). Stromal morphometry did not predict response to FUFA chemotherapy. CONCLUSIONS: Simple digital morphometry applied to a single representative H&E section identifies CRC patients with a >50% higher risk of disease recurrence. This technique can reliably partition patients into subpopulations with different risks of tumour recurrence in a simple and cost-effective manner. Further prospective validation is warranted.

Comprehensive genomic meta-analysis identifies intra-tumoural stroma as a predictor of survival in patients with gastric cancer.

OBJECTIVE: Gastric adenocarcinoma (gastric cancer, GC) is a major cause of global cancer mortality. Identifying molecular programmes contributing to GC patient survival may improve our understanding of GC pathogenesis, highlight new prognostic factors and reveal novel therapeutic targets. The authors aimed to produce a comprehensive inventory of gene expression programmes expressed in primary GCs, and to identify those expression programmes significantly associated with patient survival. DESIGN: Using a network-modelling approach, the authors performed a large-scale meta-analysis of GC transcriptome data integrating 940 gastric transcriptomes from multiple independent patient cohorts. The authors analysed a training set of 428 GCs and 163 non-malignant gastric samples, and a validation set of 288 GCs and 61 non-malignant gastric samples. RESULTS: The authors identified 178 gene expression programmes ('modules') expressed in primary GCs, which were associated with distinct biological processes, chromosomal location patterns, cis-regulatory motifs and clinicopathological parameters. Expression of a transforming growth factor ß (TGF-ß) signalling associated 'super-module' of stroma-related genes consistently predicted patient survival in multiple GC validation cohorts. The proportion of intra-tumoural stroma, quantified by morphometry in tissue sections from gastrectomy specimens, was also significantly associated with stromal super-module expression and GC patient survival. CONCLUSION: Stromal gene expression predicts GC patient survival in multiple independent cohorts, and may be closely related to the intra-tumoural stroma proportion, a specific morphological GC phenotype. These findings suggest that therapeutic approaches targeting the GC stroma may merit evaluation.

Histopathological work-up of resection specimens, local excisions and biopsies in colorectal cancer.

The pathological examination of material removed from patients with colorectal neoplasia is important. It provides a wide range of information on, for example, the quality and completeness of excision, the stage and biological aggressiveness, the need for further therapy, and response to therapy. Molecular testing adds valuable information on genetic risk and is required before treatment with anti-EGF-r antibodies. This article highlights the value derived from macroscopic inspection of surgical specimens, careful microscopy and excellent reporting according to national guidelines. Increasing use of a number of preoperative therapies and combinations in rectal cancer change the pathological features found and a standardised approach to the diagnosis of complete response is required. It touches upon the issues with frequent changes in TNM staging and the difficulties these changes are causing. The widespread introduction of bowel cancer screening is changing the stage of presentation of colorectal cancer leading to increasing numbers of local excisions and polyp cancers.