Loss Function for Learning Non-annotated Detection Targets: Evaluation on Liver Tumor Detection from Abdominal Ultrasound Images

Abstract

Collecting accurately annotated training data in the medical field is challenging. This paper proposes a method to adapt the YOLOv3 loss function for detecting metastatic liver cancer from abdominal ultrasound images, even with incomplete annotations. We aim to enhance detection accuracy for unannotated tumors by incorporating tumor-type-specific weights into the background component of the YOLOv3 loss function. After conducting experiments where only the weight for metastatic liver cancer was reduced, we observed improved recall rates for this tumor type. Furthermore, when the weight was set to 0.1, the precision for annotated metastatic liver cancer decreased by approximately 10％, due to the detection of non-annotated tumors during evaluation. It is important to note that the detection accuracy for other tumor types did not diminish. These findings indicate the effectiveness of our proposed method.