Surpassing the Human Accuracy: Detecting Gallbladder Cancer from USG Images with Curriculum Learning

We explore the potential of CNN-based models for gallbladder cancer (GBC)detection from ultrasound (USG) images as no prior study is known. USG is themost common diagnostic modality for GB diseases due to its low cost andaccessibility. However, USG images are challenging to analyze due to low imagequality, noise, and varying viewpoints due to the handheld nature of thesensor. Our exhaustive study of state-of-the-art (SOTA) image classificationtechniques for the problem reveals that they often fail to learn the salient GBregion due to the presence of shadows in the USG images. SOTA object detectiontechniques also achieve low accuracy because of spurious textures due to noiseor adjacent organs. We propose GBCNet to tackle the challenges in our problem.GBCNet first extracts the regions of interest (ROIs) by detecting the GB (andnot the cancer), and then uses a new multi-scale, second-order poolingarchitecture specializing in classifying GBC. To effectively handle spurioustextures, we propose a curriculum inspired by human visual acuity, whichreduces the texture biases in GBCNet. Experimental results demonstrate thatGBCNet significantly outperforms SOTA CNN models, as well as the expertradiologists. Our technical innovations are generic to other USG image analysistasks as well. Hence, as a validation, we also show the efficacy of GBCNet indetecting breast cancer from USG images. Project page with source code, trainedmodels, and data is available at https://gbc-iitd.github.io/gbcnet