Predicting cancer outcomes from whole slide images via hybrid supervision learning

Abstract

Collaboratively leveraging limited pixel-level segmentation annotations and large-scale slide-level classification labels in hybrid supervision learning can significantly enhance model performance. However, the direct application of this approach within computational pathology presents challenges as end-to-end classification models grapple with processing high-resolution whole slide images (WSIs). An alternative approach is to use patch-based models with pixel-level pseudo-labels, but these models can be susceptible to the cumulative effects of noisy labels, leading to convergence and drift problems during iterative training. To surmount these hurdles, we propose a hybrid supervision learning framework tailored for pathological image classification. Our method employs coarse classification labels to optimize pixel-level pseudo-labels and incorporates a comprehensive strategy to diminish false negatives and positives throughout the segmentation process. This framework holistically integrates the supervised information derived from segmentation and classification procedures, and is applicable to the general classification of high-resolution images, thereby boosting both specificity and sensitivity. We assess our proposed method’s effectiveness using one publicly accessible dataset and two proprietary datasets, collectively constituting over 10,000 pathological images of various disease types such as gastric, cervical, and breast cancer. Our experimental results reveal a 100% sensitivity rate in slide-level classification tasks, simultaneously reducing the false-positive rate to a mere third of the state-of-the-art. In conclusion, this paper presents a potent instrument for the precise and efficient classification of high-resolution pathological images, with promising results showcased across a wide array of datasets and disease types. Code is available at https://github.com/JarveeLee/HybridSupervisionLearning_Pathology.