Multi-Task Weak Supervision Enables Automated Abnormality Localization in Whole-Body FDG-PET

The availability of large, labeled datasets has fueled recent progress in machine learning for med-14 ical imaging. Breakthrough studies in chest radiograph diagnosis, skin and mammographic lesion 15 classification, and diabetic retinopathy detection have relied on hundreds of thousands of labeled 16 training examples [1–4]. However, for many diagnostic interpretation tasks, labeled datasets of this 17 size are not readily available either because (1) the task includes rare diagnoses for which training 18 examples are hard to find, and/or (2) the diagnoses are not recorded in a structured way within elec-19 tronic medical records, requiring physicians to manually reinterpret exams or extract labels from 20 free-text reports. Even if datasets are manually annotated, common changes in the underlying data 21 distribution (eg scanner type, imaging protocol, post-processing techniques, patient population, 22 or clinical classification schema) could rapidly render the models they support obsolete. Further, it 23 is commonly the case that medical datasets are labeled using incomplete label ontologies, leading 24 to undesirable variation in performance on unlabeled subsets of the data (eg rare disease types), 25 a problem commonly referred to as hidden stratification [5]. 26