A refined weakly-supervised semantic segmentation method on lung adenocarcinoma histopathology images

Automatic segmentation of tissues from histopathology images is of greatsignificance to reduce the annotation cost hence improve the efficiency of lungadenocarcinoma diagnosis. In the field of weakly supervised semanticsegmentation (WSSS), class activation mapping (CAM) is a widely usedtechnology that only uses image-level label to locate target objects. While themethod based on CAM can locate the target object precisely, it cannot fill theentire object. Therefore, In this paper, we propose a novel method of joint refinederasure attention module (REM) and class re-activation mapping (ReCAM) forimage-level semantic segmentation of histopathology images. Specifically, REMconsists of two components: channel attention and progressive erasure attention.Channel attention dynamically adjusts the weights of each channel based on theirrelevance and progressive erasure attention selectively removes the mostdiscriminative regions to encourage the model to explore more features. Thereason for this is to keep the recognition ability of each feature while expandingthe seed region. Moreover, due to the reason that different tissues may includesimilar features caused by aggressive nature of the malignancy, we employ thereactivation method to enhance the model’s discriminative power. This methodfine-tunes the model parameters by learning an additional fully connected layerwith softmax cross-entropy loss (SCE). We evaluate the proposed model on therecently released LUAD-HistoSeg dataset. The results illustrate that our modelachieved a 76.53% MIoU and outperforms state-of-the-art weakly-supervisedlearning methods.