Whole-body low-dose computed tomography in patients with newly diagnosed multiple myeloma predicts cytogenetic risk: a deep learning radiogenomics study.

OBJECTIVE: To develop a whole-body low-dose CT (WBLDCT) deep learning model and determine its accuracy in predicting the presence of cytogenetic abnormalities in multiple myeloma (MM). MATERIALS AND METHODS: WBLDCTs of MM patients performed within a year of diagnosis were included. Cytogenetic assessments of clonal plasma cells via fluorescent in situ hybridization (FISH) were used to risk-stratify patients as high-risk (HR) or standard-risk (SR). Presence of any of del(17p), t(14;16), t(4;14), and t(14;20) on FISH was defined as HR. The dataset was evenly divided into five groups (folds) at the individual patient level for model training. Mean and standard deviation (SD) of the area under the receiver operating curve (AUROC) across the folds were recorded. RESULTS: One hundred fifty-one patients with MM were included in the study. The model performed best for t(4;14), mean (SD) AUROC of 0.874 (0.073). The lowest AUROC was observed for trisomies: AUROC of 0.717 (0.058). Two- and 5-year survival rates for HR cytogenetics were 87% and 71%, respectively, compared to 91% and 79% for SR cytogenetics. Survival predictions by the WBLDCT deep learning model revealed 2- and 5-year survival rates for patients with HR cytogenetics as 87% and 71%, respectively, compared to 92% and 81% for SR cytogenetics. CONCLUSION: A deep learning model trained on WBLDCT scans predicted the presence of cytogenetic abnormalities used for risk stratification in MM. Assessment of the model's performance revealed good to excellent classification of the various cytogenetic abnormalities.

THA-AID: Deep Learning Tool for Total Hip Arthroplasty Automatic Implant Detection With Uncertainty and Outlier Quantification.

BACKGROUND: Revision total hip arthroplasty (THA) requires preoperatively identifying in situ implants, a time-consuming and sometimes unachievable task. Although deep learning (DL) tools have been attempted to automate this process, existing approaches are limited by classifying few femoral and zero acetabular components, only classify on anterior-posterior (AP) radiographs, and do not report prediction uncertainty or flag outlier data. METHODS: This study introduces Total Hip Arhtroplasty Automated Implant Detector (THA-AID), a DL tool trained on 241,419 radiographs that identifies common designs of 20 femoral and 8 acetabular components from AP, lateral, or oblique views and reports prediction uncertainty using conformal prediction and outlier detection using a custom framework. We evaluated THA-AID using internal, external, and out-of-domain test sets and compared its performance with human experts. RESULTS: THA-AID achieved internal test set accuracies of 98.9% for both femoral and acetabular components with no significant differences based on radiographic view. The femoral classifier also achieved 97.0% accuracy on the external test set. Adding conformal prediction increased true label prediction by 0.1% for acetabular and 0.7 to 0.9% for femoral components. More than 99% of out-of-domain and >89% of in-domain outlier data were correctly identified by THA-AID. CONCLUSIONS: The THA-AID is an automated tool for implant identification from radiographs with exceptional performance on internal and external test sets and no decrement in performance based on radiographic view. Importantly, this is the first study in orthopedics to our knowledge including uncertainty quantification and outlier detection of a DL model.

Quantifying Uncertainty in Deep Learning of Radiologic Images.

In recent years, deep learning (DL) has shown impressive performance in radiologic image analysis. However, for a DL model to be useful in a real-world setting, its confidence in a prediction must also be known. Each DL model's output has an estimated probability, and these estimated probabilities are not always reliable. Uncertainty represents the trustworthiness (validity) of estimated probabilities. The higher the uncertainty, the lower the validity. Uncertainty quantification (UQ) methods determine the uncertainty level of each prediction. Predictions made without UQ methods are generally not trustworthy. By implementing UQ in medical DL models, users can be alerted when a model does not have enough information to make a confident decision. Consequently, a medical expert could reevaluate the uncertain cases, which would eventually lead to gaining more trust when using a model. This review focuses on recent trends using UQ methods in DL radiologic image analysis within a conceptual framework. Also discussed in this review are potential applications, challenges, and future directions of UQ in DL radiologic image analysis.

A deep learning algorithm for detecting lytic bone lesions of multiple myeloma on CT.

BACKGROUND: Whole-body low-dose CT is the recommended initial imaging modality to evaluate bone destruction as a result of multiple myeloma. Accurate interpretation of these scans to detect small lytic bone lesions is time intensive. A functional deep learning) algorithm to detect lytic lesions on CTs could improve the value of these CTs for myeloma imaging. Our objectives were to develop a DL algorithm and determine its performance at detecting lytic lesions of multiple myeloma. METHODS: Axial slices (2-mm section thickness) from whole-body low-dose CT scans of subjects with biochemically confirmed plasma cell dyscrasias were included in the study. Data were split into train and test sets at the patient level targeting a 90%/10% split. Two musculoskeletal radiologists annotated lytic lesions on the images with bounding boxes. Subsequently, we developed a two-step deep learning model comprising bone segmentation followed by lesion detection. Unet and "You Look Only Once" (YOLO) models were used as bone segmentation and lesion detection algorithms, respectively. Diagnostic performance was determined using the area under the receiver operating characteristic curve (AUROC). RESULTS: Forty whole-body low-dose CTs from 40 subjects yielded 2193 image slices. A total of 5640 lytic lesions were annotated. The two-step model achieved a sensitivity of 91.6% and a specificity of 84.6%. Lesion detection AUROC was 90.4%. CONCLUSION: We developed a deep learning model that detects lytic bone lesions of multiple myeloma on whole-body low-dose CTs with high performance. External validation is required prior to widespread adoption in clinical practice.

A Comparison of Three Different Deep Learning-Based Models to Predict the MGMT Promoter Methylation Status in Glioblastoma Using Brain MRI.

Glioblastoma (GBM) is the most common primary malignant brain tumor in adults. The standard treatment for GBM consists of surgical resection followed by concurrent chemoradiotherapy and adjuvant temozolomide. O-6-methylguanine-DNA methyltransferase (MGMT) promoter methylation status is an important prognostic biomarker that predicts the response to temozolomide and guides treatment decisions. At present, the only reliable way to determine MGMT promoter methylation status is through the analysis of tumor tissues. Considering the complications of the tissue-based methods, an imaging-based approach is preferred. This study aimed to compare three different deep learning-based approaches for predicting MGMT promoter methylation status. We obtained 576 T2WI with their corresponding tumor masks, and MGMT promoter methylation status from, The Brain Tumor Segmentation (BraTS) 2021 datasets. We developed three different models: voxel-wise, slice-wise, and whole-brain. For voxel-wise classification, methylated and unmethylated MGMT tumor masks were made into 1 and 2 with 0 background, respectively. We converted each T2WI into 32 × 32 × 32 patches. We trained a 3D-Vnet model for tumor segmentation. After inference, we constructed the whole brain volume based on the patch's coordinates. The final prediction of MGMT methylation status was made by majority voting between the predicted voxel values of the biggest connected component. For slice-wise classification, we trained an object detection model for tumor detection and MGMT methylation status prediction, then for final prediction, we used majority voting. For the whole-brain approach, we trained a 3D Densenet121 for prediction. Whole-brain, slice-wise, and voxel-wise, accuracy was 65.42% (SD 3.97%), 61.37% (SD 1.48%), and 56.84% (SD 4.38%), respectively.

Reproducibility of Deep Learning Algorithms Developed for Medical Imaging Analysis: A Systematic Review.

Since 2000, there have been more than 8000 publications on radiology artificial intelligence (AI). AI breakthroughs allow complex tasks to be automated and even performed beyond human capabilities. However, the lack of details on the methods and algorithm code undercuts its scientific value. Many science subfields have recently faced a reproducibility crisis, eroding trust in processes and results, and influencing the rise in retractions of scientific papers. For the same reasons, conducting research in deep learning (DL) also requires reproducibility. Although several valuable manuscript checklists for AI in medical imaging exist, they are not focused specifically on reproducibility. In this study, we conducted a systematic review of recently published papers in the field of DL to evaluate if the description of their methodology could allow the reproducibility of their findings. We focused on the Journal of Digital Imaging (JDI), a specialized journal that publishes papers on AI and medical imaging. We used the keyword "Deep Learning" and collected the articles published between January 2020 and January 2022. We screened all the articles and included the ones which reported the development of a DL tool in medical imaging. We extracted the reported details about the dataset, data handling steps, data splitting, model details, and performance metrics of each included article. We found 148 articles. Eighty were included after screening for articles that reported developing a DL model for medical image analysis. Five studies have made their code publicly available, and 35 studies have utilized publicly available datasets. We provided figures to show the ratio and absolute count of reported items from included studies. According to our cross-sectional study, in JDI publications on DL in medical imaging, authors infrequently report the key elements of their study to make it reproducible.

Development of a deep learning model for the histologic diagnosis of dysplasia in Barrett's esophagus.

BACKGROUND AND AIMS: The risk of progression in Barrett's esophagus (BE) increases with development of dysplasia. There is a critical need to improve the diagnosis of BE dysplasia, given substantial interobserver disagreement among expert pathologists and overdiagnosis of dysplasia by community pathologists. We developed a deep learning model to predict dysplasia grade on whole-slide imaging. METHODS: We digitized nondysplastic BE (NDBE), low-grade dysplasia (LGD), and high-grade dysplasia (HGD) histology slides. Two expert pathologists confirmed all histology and digitally annotated areas of dysplasia. Training, validation, and test sets were created (by a random 70/20/10 split). We used an ensemble approach combining a "you only look once" model to identify regions of interest and histology class (NDBE, LGD, or HGD) followed by a ResNet101 model pretrained on ImageNet applied to the regions of interest. Diagnostic performance was determined for the whole slide. RESULTS: We included slides from 542 patients (164 NDBE, 226 LGD, and 152 HGD) yielding 8596 bounding boxes in the training set, 1946 bounding boxes in the validation set, and 840 boxes in the test set. When the ensemble model was used, sensitivity and specificity for LGD was 81.3% and 100%, respectively, and >90% for NDBE and HGD. The overall positive predictive value and sensitivity metric (calculated as F1 score) was .91 for NDBE, .90 for LGD, and 1.0 for HGD. CONCLUSIONS: We successfully trained and validated a deep learning model to accurately identify dysplasia on whole-slide images. This model can potentially help improve the histologic diagnosis of BE dysplasia and the appropriate application of endoscopic therapy.

A deep learning model for discriminating true progression from pseudoprogression in glioblastoma patients.

INTRODUCTION: Glioblastomas (GBMs) are highly aggressive tumors. A common clinical challenge after standard of care treatment is differentiating tumor progression from treatment-related changes, also known as pseudoprogression (PsP). Usually, PsP resolves or stabilizes without further treatment or a course of steroids, whereas true progression (TP) requires more aggressive management. Differentiating PsP from TP will affect the patient's outcome. This study investigated using deep learning to distinguish PsP MRI features from progressive disease. METHOD: We included GBM patients with a new or increasingly enhancing lesion within the original radiation field. We labeled those who subsequently were stable or improved on imaging and clinically as PsP and those with clinical and imaging deterioration as TP. A subset of subjects underwent a second resection. We labeled these subjects as PsP, or TP based on the histological diagnosis. We coregistered contrast-enhanced T1 MRIs with T2-weighted images for each patient and used them as input to a 3-D Densenet121 model and using five-fold cross-validation to predict TP vs PsP. RESULT: We included 124 patients who met the criteria, and of those, 63 were PsP and 61 were TP. We trained a deep learning model that achieved 76.4% (range 70-84%, SD 5.122) mean accuracy over the 5 folds, 0.7560 (range 0.6553-0.8535, SD 0.069) mean AUROCC, 88.72% (SD 6.86) mean sensitivity, and 62.05% (SD 9.11) mean specificity. CONCLUSION: We report the development of a deep learning model that distinguishes PsP from TP in GBM patients treated per the Stupp protocol. Further refinement and external validation are required prior to widespread adoption in clinical practice.

Abstract screening using the automated tool Rayyan: results of effectiveness in three diagnostic test accuracy systematic reviews.

OBJECTIVE: To evaluate the performance of the automated abstract screening tool Rayyan. METHODS: The records obtained from the search for three systematic reviews were manually screened in four stages. At the end of each stage, Rayyan was used to predict the eligibility score for the remaining records. At two different thresholds (≤2.5 and < 2.5 for exclusion of a record) Rayyan-generated ratings were compared with the decisions made by human reviewers in the manual screening process and the tool's accuracy metrics were calculated. RESULTS: Two thousand fifty-four records were screened manually, of which 379 were judged to be eligible for full-text assessment, and 112 were eventually included in the final review. For finding records eligible for full-text assessment, at the threshold of < 2.5 for exclusion, Rayyan managed to achieve sensitivity values of 97-99% with specificity values of 19-58%, while at the threshold of ≤2.5 for exclusion it had a specificity of 100% with sensitivity values of 1-29%. For the task of finding eligible reports for inclusion in the final review, almost similar results were obtained. DISCUSSION: At the threshold of < 2.5 for exclusion, Rayyan managed to be a reliable tool for excluding ineligible records, but it was not much reliable for finding eligible records. We emphasize that this study was conducted on diagnostic test accuracy reviews, which are more difficult to screen due to inconsistent terminology.

A Multi-View Deep Learning Model for Thyroid Nodules Detection and Characterization in Ultrasound Imaging.

Thyroid Ultrasound (US) is the primary method to evaluate thyroid nodules. Deep learning (DL) has been playing a significant role in evaluating thyroid cancer. We propose a DL-based pipeline to detect and classify thyroid nodules into benign or malignant groups relying on two views of US imaging. Transverse and longitudinal US images of thyroid nodules from 983 patients were collected retrospectively. Eighty-one cases were held out as a testing set, and the rest of the data were used in five-fold cross-validation (CV). Two You Look Only Once (YOLO) v5 models were trained to detect nodules and classify them. For each view, five models were developed during the CV, which was ensembled by using non-max suppression (NMS) to boost their collective generalizability. An extreme gradient boosting (XGBoost) model was trained on the outputs of the ensembled models for both views to yield a final prediction of malignancy for each nodule. The test set was evaluated by an expert radiologist using the American College of Radiology Thyroid Imaging Reporting and Data System (ACR-TIRADS). The ensemble models for each view achieved a mAP0.5 of 0.797 (transverse) and 0.716 (longitudinal). The whole pipeline reached an AUROC of 0.84 (CI 95%: 0.75-0.91) with sensitivity and specificity of 84% and 63%, respectively, while the ACR-TIRADS evaluation of the same set had a sensitivity of 76% and specificity of 34% (p-value = 0.003). Our proposed work demonstrated the potential possibility of a deep learning model to achieve diagnostic performance for thyroid nodule evaluation.