Conventional size object detection has been extensively studied, whereas researches concerning ultrasmall object detection are rare due to lack of dataset. Here, considering that the stapes in the ear is the smallest bone in our body, we have collected the largest stapedial otosclerosis detection dataset from 633 stapedial otosclerosis patients and 269 normal cases to promote this direction. Nevertheless, noisy classification labels in our dataset are inevitable due to various subjective and objective factors, and this situation prevails in various fields. In this paper, we propose a novel and general noise tolerant loss function named Adaptive Cross Entropy (ACE) which needs no fine-tuning of hyperparameters for training with noisy labels. We provide both theoretical and empirical analyses for the proposed ACE loss and demonstrate its effectiveness in multiple public datasets. Besides, we find high-resolution representations crucial for ultrasmall object detection and present an auxiliary backbone called W-Net to address it accordingly. Extensive experiments demonstrate that the proposed ACE loss is able to boost the diagnosis performance under noisy label setting by a large margin. Furthermore, our W-Net can help extract sufficient high-resolution representations specialized for ultrasmall objects and achieve even better results. Hopefully, our work could provide more clues for future research on ultrasmall object detection and learning with noisy labels.
Keywords: Deep learning; Label noise; Medical image diagnosis; Object detection; Otosclerosis.
Copyright © 2022 Elsevier Ltd. All rights reserved.