Manuscript received March 31, 2026; revised April 22, 2026; accepted June 5, 2026; published October 19, 2026.

Abstract—This paper proposes a unified semi-supervised multi-task framework, Dual-Framework Cross Pseudo Supervision (DFCPS), that simultaneously performs segmentation and classification within a single shared-backbone network. Our method uses Cross Pseudo Supervision (CPS) with a confidence threshold of 0.95, updating every three training steps, and applies class-weighted losses to handle a severe 229:1 class imbalance. This design cuts computational cost by half compared to using separate models. On the Hyper-Kvasir dataset (11,662 images, 23 disease classes), our model achieves 68.37% mean Intersection over Union (mIoU) and 79.11% Dice for segmentation, and 75.49% accuracy with 57.49% balanced accuracy for classification. Our ablation studies show that removing Atrous Spatial Pyramid Pooling (ASPP) drops mIoU by 2.27%, removing class weights widens the imbalance gap by 1.33%, and running CPS at every step reduces mIoU by 5.26% while increasing training time by 62%. These results show our framework works well, but also show that class imbalance remains a tough open problem.

Keywords—endoscopic imaging, gastrointestinal lesion analysis, semi-supervised learning, multi-task learning, cross pseudo supervision, class imbalance
 

Cite: Huda Dham Bader, Ali Rakan Hasan Al-Jader, and Mohammed Sabah Jarjees, "Semi-Supervised Multi-task Learning with Cross Pseudo Supervision for Gastrointestinal Lesion Analysis," Journal of Image and Graphics, Vol. 14, No. 5, pp. 831-848, 2026.

Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).