Fused Multilayer Layer-CAM Fine-Grained Spatial Feature Supervision for Surgical Phase Classification Using CNNs

Abstract

AbstractIn this paper, we propose a novel spatial context aware combined loss function to be used along with an end to end Encoder-Decoder training methodology for the task of surgical phase classification on laparoscopic cholecystectomy surgical videos. Proposed spatial context aware combined loss function leverages on the fine-grained class activation maps obtained from fused multilayer Layer-CAM for supervising the learning of surgical phase classifier. We report peak surgical phase classification accuracy of 91.95%, precision of 86.19% and recall of 83.75% on publicly available Cholec80 dataset consisting of 7 surgical phases. Our proposed method utilizes just 77% of the total number of parameters in comparison with state of the art methodology and achieves 3.4% improvement in terms of accuracy, 4.6% improvement in terms of precision and comparable recall.KeywordsLaparoscopic cholecystectomySurgical work flow analysisTransfer learningClass activation mapsCNN