Enhanced Pooling for Weakly Supervised Gigapixel WSI Training Improves Classification and Lesion Localization

A robust artificial intelligence-assisted workflow for tumor assessment in pathology requires not only accurate classification but also precise lesion localization. While current weakly supervised learning methods significantly reduce the need for extensive annotations and leverage large quantities of annotation-free whole-slide images (WSIs) to enhance classification robustness, they often fall short in segmentation accuracy. We attribute this limitation to the optimization goals in classification, which tend to focus solely on the most representative features--an approach that is particularly inefficient for WSIs with gigapixel resolution. To address this challenge, we introduce a novel approach based on streaming convolution, an end-to-end method for WSI training. Our contributions include the Rectified LogSumExp (ReLSE) pooling method and adaptive pseudo annotation generation for self-training, both designed to encourage models to learn from sub-representative features. Using only slide-level annotations from the CAMELYON16 dataset, our method achieves a significant improvement in metastasis localization, with a patch-level recall from 53.88% to 79.18% at a precision of 80%. This conclusion also holds for a dataset collected from Taipei Veterans General Hospital (TVGH), used in the assessment of lung cancer lymph node metastasis with a recall improved from 42.98% to 72.27%. The proposed method can also be integrated with additional manual annotations. Experiment results showed the semi-supervised model (recall: 91.12%) could even outperform the strongly supervised model (recall: 81.80%) on the TVGH dataset.