Revisiting the supervision level in semi-supervised learning for automated tumor segmentation: application to lymphoma FDG PET imaging

Convolutional neural networks have shown the power to segment FDG-avid tumors from PET scans. However, the training step of most of these approaches is supervised, demanding annotated data for training. To reduce this need, we previously evaluated the performance of 3D U-Net for lesion segmentation by two semi-supervised approaches using the combined loss functions of Robust FCM (RFCM) and Mumford-Shah (MS) losses for unsupervised term, and Dice and Labeled FCM (LFCM) for supervised term. In this work, we applied two combined losses (RFCM+αLFCM and MS+αDice) to consider the effect of supervision level (α) on segmentation performance on a multi-center (BC&SM) dataset. We applied two experiments utilizing the PET images of 292 patients. In experiment I, training and test data are from diffuse large B-cell (DLBCL) cases mostly from center BC. In experiment II, we applied a more realistic scenario in which the segmentation model was trained on DLBCL and primary mediastinal B-cell (PMBCL) data mostly from BC and tested on SM data. Our results showed that increasing the impact of the supervised term leads to different segmentation performances in both experiments. The best Dice score for MS-based semi-supervised approach were 0.60±0.17 (α=0.7) and 0.6±0.08 (α=0.2) for experiments I and II respectively. In FCM-based approach the best Dice score were 0.71±0.09 (α=0.7) and 0.69±0.03 (α=0.3) for experiments I and II, respectively. We concluded when test data from a center has low contribution in training data, increasing the supervision level results in lower segmentation performance. The performance drops higher by increasing α in MS-based semi-supervised approach.