Generating Clinically Translatable AI Models for Cancer Diagnostics

Abstract

Given the large heterogeneity in the oncogenesis, pathophysiology and therapeutic targeting of cancers and the significant morbidity that cancer imposes on society, standard-of-care oncotherapy has focused on earlier screening and detection, and subsequent preferential removal of lesions via a combination of chemo-, radio- and/or targeted therapy to improve overall survival. Despite significant advancements in the therapeutic arm for most cancers, the clinical diagnostic gold standard remains the biopsy, particularly for identifying the mutational status of therapeutically relevant molecular markers and oncogenic drivers, and for shedding light into the tumor microenvironment. However, biopsies are invasive (requiring organ penetration), expensive (both in terms of cost and time) and frequently has poor sensitivity. Therefore, there is a critical need for minimally-invasive, well-validated diagnostic biomarkers that can supplant biopsies and facilitate precision medicine.

The use of artificial intelligence (AI) and deep learning (DL) models trained on routinely collected imaging (e.g., CT, MRI, colposcopy) has recently emerged as a minimally-invasive alternative diagnostic biomarker in several clinical domains, with optimized models reporting near-clinician-level performance. However, translation of DL models from bench to bedside remain sparse. To be clinically translatable, deep neural networks (DNN) should be robust, computationally efficient, low-cost, and blend well with existing clinical workflows, ensuring the inputs/outputs of the model and the task it performs are most relevant to the clinician for a given use case. This is often not the case with current state-of-the-art (SOA) models, which are frequently hindered by several key methodological flaws in their design, thereby undermining their validity, and hindering clinical translation. In the context of my work, model robustness refers to two key attributes: 1. repeatability or reproducibility, defined as the ability of a model to generate near-identical predictions for the same patient under identical conditions, ensuring that the model produces precise, reliable and consistent outputs in the clinical setting; and 2. generalizability or portability, defined as the ability of a model to adapt well to domain expansion or, alternatively, the ability of a model to perform well on datasets that are out of distribution i.e., having different characteristics from the training data (e.g., different device, geography, quality, and/or patient population). There is a paucity of work in the current literature that assess one or both of these attributes, with models tending to overfit the training data distribution.

In my thesis work, I address both these attributes head-on, via comprehensively-designed, biologically-inspired AI pipelines that are optimized specifically for clinical translation. Key technical innovations highlighted in my work include repeatability-centric model optimization and combination loss functions with Monte Carlo dropout (for improved repeatability), as well as novel metrics for distribution distance characterization and optimized retraining (for improved generalizability). My work incorporates these innovations into DL-based pipelines in several oncology domains. In particular, for cervical cancer, we developed a multi-stage AI pipeline utilizing a cervical-colposcope-thermocoagulator device in a workflow involving, in sequence: 1. image capture, 2. cervix detection, 3. an image quality classifier that filters the images for quality, 4. a diagnostic classifier that classifies the cervix into one of three diagnoses (normal, intermediate and precancer or above), and 5. thermal ablation using the attached thermocoagulator if indicated, otherwise surgical excision or deferral. Chapters 1 through 4 highlight multi-stage, comprehensive development of our cervical cancer screening pipeline, together with the technical innovations explored in our work to assess model repeatability and generalizability. Given the high clinical burden of cervical cancer, this pipeline can serve as a triage tool for human papillomavirus (HPV) positive women in low resource settings, where existing screening methods, such as biopsy and visual inspection after application of acetic acid (VIA) are invasive, expensive and/or inaccurate. A large-scale prospective efficacy and effectiveness study of this pipeline is already underway at the partner institutions of our consortium in low and middle-income countries. Further work towards generating AI models in other clinical domains, in particular, brain tumors is highlighted in the Appendix Chapter 5.