Ali Bashashati Saghezchi

The successful integration of deep learning in medical imaging relies upon the reliability and predictiveness of the models. It is important that these models provide accurate predictions for known classes while also delivering well-calibrated uncertainty estimates, especially for unseen classes and anomalies that are regarded as the out-of-distribution (OOD) data, distinct from the data used for model training. Accurate uncertainty estimates can potentially reduce the adverse effects of OOD regions on the target classification task in a clinical workflow. This, in turn, can prevent models from silently failing when confronted with unfamiliar diseases or abnormalities. Our work introduces two distinct approaches, namely M-Branch (Multi-Branch) and VPS (Virtual Patch Synthesis), for training multi-instance learning (MIL) models in histopathology, endowing them with the capability to effectively estimate predictive uncertainty. We conduct a comprehensive performance evaluation by comparing our proposed models to a state-of-the-art MIL model in whole-slide image (WSI) classification, equipped with temperature scaling for enhanced calibration, referred to as CLAM-T, focusing on the task of OOD detection.In our study, we consider the classification of Non-Small Cell Lung Cancer (NSCLC) subtypes, primarily distinguishing between LUAD (lung adenocarcinoma) and LUSC (lung squamous cell carcinoma) as in-distribution classes, while also differentiating NSCLC as in-distribution from Lower Grade Glioma (LGG) as out-of-distribution. Our top-performing model, M-Branch, efficiently estimates predictive uncertainty through the deployment of multiple branches of attention-based networks, complemented by a diversity-promoting loss. We employ two key evaluation metrics, FPR95 and AUC, to assess OOD detection performance. M-Branch excels in this regard, achieving an FPR95 of 38.39 and an AUC of 84.86, outperforming both VPS (FPR95: 49.76, AUC: 81.95) and CLAM-T (FPR95: 49.29, AUC: 83.00). Moreover, we demonstrate that the incorporation of a meta-loss function within M-Branch significantly enhances OOD detection, as evident from the improvements in FPR95 and AUC.Our research makes a substantial contribution to the field of medical image analysis by equipping MIL models with the ability to estimate predictive uncertainty effectively. These advancements have promising implications for enhancing the reliability and performance of deep learning models in medical imaging and digital pathology, particularly in real-world healthcare applications.

View record

Pancreatic ductal adenocarcinoma (PDAC) mortality rates are projected to rise by 2030 due to factors such as delayed diagnosis and resistance to chemotherapy and radiation therapy. A key challenge in treating PDAC is the lack of biomarkers for predicting treatment effectiveness and chemotherapy resistance. Researchers suggest a binary subtype system, basal-like and classical, can predict treatment selection and response, but identifying these subtypes requires costly and time-consuming RNA profiling.Histopathology, which provides an inexpensive and convenient visual readout of disease biology, has been essential in cancer diagnosis and prognosis for over a century. Artificial intelligence (AI) has recently been successfully applied to histopathology data, with AI-based models potentially outperforming traditional pathology assessments. However, an AI expert is needed to utilize and interpret these techniques.This research aimed to: 1) develop an AI-based pipeline to identify and detect histological features for classifying PDAC molecular subtypes, and 2) generalize the pipeline using a “Machine Learning Workflow Engine” and a “Web-based Slide Manager and Annotator” for processing and interpreting histopathology data.The researchers used the developed infrastructures to train and evaluate a deep-learning model for classifying PDAC patients into prognostic subgroups. They used 130 histological slides from the TCGA-PAAD dataset for training and 81 slides from 19 patients from an in-house dataset as the external test dataset. A two-step machine learning model was trained: 1) a classifier distinguishing tumor patches from stroma patches, and 2) a classifier predicting the molecular subtype of a slide based on tumor patches. The tumor/stroma classifier showed excellent performance with an AUC of 96.18% ± 1.84%, while the subtype classifier achieved a balanced accuracy of 96.19% ± 2.45% at the slide level. The model correctly classified 83.03% ± 6.35 of the patients' tumor molecular subtypes in the validation cohort. This classifier is the first to categorize PDAC patients based on biopsy samples.

View record

Neuroendocrine neoplasms (NENs) are uncommon neoplasms, which can arise in cells across the body. The last few decades have seen NEN incidence increase by 7-fold. NEN grades are morphologically indistinguishable for well-differentiated tumours and are determined from the tumour proliferation. Cell proliferation is defined by the number of mitotic figures in a 2mm² area for H&E slides and by the percent of Ki-67 positive cells in Ki-67 slides. Unfortunately, these measures suffer inter- and intra-observer variability, and are cumbersome to quantify. We developed a novel machine learning framework to identify candidate mitotic figures in H&E images, calculate the Ki-67 index in Ki-67 images, and aggregate proliferating features to grade NENs. Our work included 186 gastroenteropancreatic NENs with 385 samples from across British Columbia with two different stains (247 H&E and 138 Ki-67 images) and patient-level labels for grade. The H&E portion of the framework achieved a balanced accuracy of 72.1% across 6-folds for three-class classification (G1, G2, G3). We demonstrated that the survival outcomes for computer assigned grades are comparable to those based on pathologist assessment, with c-index values of 0.63 and 0.64, respectively. Our Ki-67 algorithm achieves a balanced accuracy of 83.9% to predict the grade of Ki-67 slides. With Ki-67 added to our pipeline, grading improves the balanced accuracy to 78.5%. Analysis of the survival outcome for pathologist assessed G1s demonstrated significant (p-value=0.01) separation amongst samples the algorithm assigned to a higher grade (n=27; median survival 4.63 years) compared to concordant G1 samples (n=55; median survival 10.13 years). By identifying both mitotic figures and Ki-67 positive cells, our method may provide a tool to verify hot-spots in both H&E and Ki-67 slides for further analysis of grades. Misclassified G1 patients have lower median survival, and further analysis is needed to determine if this group should be considered a different clinical entity and if our machine learning approach has identified two distinct populations within the grade 1 group.

View record

Most current deep learning models for hematoxylin and eosin (H&E) histopathologyimage analysis lack the power of generalization to datasets collected from otherinstitutes due to the domain shift in the data. While graph convolutional neural networkshave shown significant potential in natural and histopathology images, theiruse in histopathology images has only been studied using a single magnification ormulti-magnification with late fusion.In this thesis, we study the domain shift problem with multiple instance learning(MIL) on prostate cancer datasets collected from different centers.First, we develop a novel center-based H&E color augmentation for cross-centermodel generalization. While previous work used methods such as random augmentation,color normalization, or learning domain-independent features to improvethe robustness of the model to changes in H&E stains, our method first augmentsthe H&E color space of the source dataset to color space of both datasets and thenadds random color augmentation. Our method covers the larger range of the colordistribution of both institutions resulting in a better generalization.Next, to leverage the multi-magnification information and early fusion with graphconvolutional networks, we handle different embedding spaces at each magnificationby introducing the Multi-Scale Relational Graph Convolutional Network (MSRGCN)as a novel MIL method. We model histopathology image patches and theirrelation with neighboring patches and patches at other magnifications as a graph.To pass the information between different magnification embedding spaces, we defineseparate message-passing neural networks based on the node and edge type.Our proposed color adaptation method improves the model performance on boththe source and target datasets, and has the best performance on the unlabeled target dataset compared to State-Of-The-Art (SOTA), showing promise as an approach tolearning generalizable features for histopathology image analysis.We also compare our MS-RGCN with multiple SOTA methods with evaluations onseveral source and held-out datasets. Our method outperforms the SOTA on all ofthe datasets and image types consisting of tissue microarrays, whole-mount slideregions, and whole-slide images. Through an ablation study, we test and show thevalue of the pertinent design features of the MS-RGCN.

View record

The full abstract for this thesis is available in the body of the thesis, and will be available when the embargo expires.

View record

Ovarian carcinoma is the deadliest cancer of the female reproductive system in North America. There are five major histological subtypes which require different treatments. Pathologists diagnose these histotypes by examining hematoxylin and eosin (H&E)-stained whole slide images (WSIs) of tissue. However, histotype diagnosis is not simple, with poor interobserver agreement between general pathologists (Cohen’s kappa 0.54-0.67). We hypothesize that latest machine learning (ML)-based image classification models may be able to recognize ovarian carcinoma histotype sufficiently well that they could aid pathologists in diagnosis.However, the color variation of H&E-stained tissues, especially those from different centers/hospitals, is a longstanding challenge for applications of AI in digital pathology. First, we investigate eight color normalization algorithms as a preprocessing step for artificial intelligence (AI)-based classification. Using multiple datasets of different cancer types, reference images, and cross-validation splits, we show that color normalization significantly improves the classification accuracy of WSIs when the train and test data are from separate institutions (ovarian cancer: 0.25 AUC increase, p = 1.6 e-05, pleural cancer: 0.21 AUC increase, p = 1.4 e-10). Furthermore, we introduce a novel augmentation strategy by mixing color-normalized images using three easily accessible algorithms that consistently improves the diagnosis of test images from external institutions.Secondly, we train four different deep convolutional neural networks to automatically classify H&E-stained images of epithelial ovarian carcinoma using the largest training dataset to date (948slides corresponding to 485 patients). Performance is assessed on an independent test set of 60 patients from another institution. The best performing model achieves a mean diagnostic concordance of 80.97% (Cohen’s kappa 0.7547). As well, in 4 of 8 cases misclassified by ML from the external dataset, two expert subspecialty pathologists rendered diagnoses, based on blind review of the WSIs, that agree with AI rather than the integrated reference diagnosis.Our results indicate that color normalization can reliably improve AI-based diagnosis of WSIs sourced from multiple centers, and specifically that an ML-based ovarian carcinoma classifier is ready for clinical validation studies as an adjunct for informing histotype diagnosis, thereby supporting histotype-specific ovarian cancer treatment and accordingly reduce the deadliness of this disease.

View record

Background: In contrast to histopathological assessment, molecular subtyping of Endometrial Cancer (EC) provides a reproducible classification system with significant prognostic value. Proactive Molecular Risk Classifier for Endometrial Cancer (ProMisE) was developed as a practical, cost-efficient, and therapeutically beneficial molecular classifier, replacing complex genomic tests. ProMisE stratifies EC into four subtypes: (1) POLE mutant, (2) Mismatch repair deficient, (3) p53 abnormal (p53abn) by immunohistochemistry, and (4) No Specific Molecular Profile (NSMP), which lacks any of the distinguishing features of the other three subtypes. Although ProMisE has provided significant prognostic value, there are clinical outliers within its four subtypes. This is especially evident in the largest ProMisE subtype, NSMP, accounting for about half of all ECs, where a fraction of patients encounter a very aggressive disease course, similar to the behavior of patients diagnosed with p53abn. Method: We considered the problem of refining the EC NSMP subtype using ubiquitous histopathology images. We hypothesized that evaluating the digital hematoxylin and eosin-stained images of NSMP could discern clinical outcome outliers. To this end, we designed an image analysis framework utilizing Artificial Intelligence (AI) to detect NSMP patients with comparable histological characteristics to the p53abn subtype. The analysis included various preprocessing steps, deep neural networks classifying the subtype of images, and survival and genomic analyses. Finding: Exploiting an AI-based methodology, we have expanded the NSMP subtype into two subgroups: ‘p53abn-like’ NSMPs and the rest of the NSMP cases. The former consists of patients diagnosed with NSMP by ProMisE, yet our AI-based analysis labeled them as p53abn due to morphological similarities. With following similar trends in two independent datasets, ‘p53abn-like’ NSMPs displayed comparable clinical behavior to p53abn, where they had markedly unfavorable outcomes in comparison with the remainder of the NSMP cases. In addition, the extensive genomic analysis suggested that ‘p53abn- like’ NSMPs had significantly higher fractions of genome altered than NSMPs in both datasets, validating our initial hypothesis in a different domain of data. We also discovered that ‘p53abn-like’ NSMPs patients might not benefit from hormone therapy. These findings emphasize the potential of AI screening as a stratification tool within ProMisE.

View record