Multitask learning for ultrasound image classification and segmentation of endometrium in the diagnosis of ectopic pregnancy

OBJECTIVE: To develop an automated tool that performs classification and segmentation of endometrium for ectopic pregnancy diagnosis before gestational sac identification through ultrasound and serial β-hCG trends are established. MATERIALS AND METHODS: The dataset comprised 1673 transvaginal ultrasonography images from early normal pregnancy (n = 382), ectopic pregnancy (n = 422), complete abortion (n = 445), and incomplete abortion (n = 424). We propose a multitask deep learning network that leverages information from various tasks, which is integrated using a weighted sum and bias, with uncertainty considerations incorporated during the training phase. Grad-CAM was employed to generate heatmaps highlighting image regions most influential in the model's decision-making process. RESULTS: The proposed model demonstrated superior performance in distinguishing ectopic pregnancy from early normal pregnancy, with an accuracy of 0.77, an AUC of 0.85, and both a sensitivity and specificity of 0.77. Additionally, images from cases of complete and incomplete abortion were incorporated into the training dataset for differentiating extrauterine pregnancy (ectopic pregnancy), with these cases categorized alongside intrauterine pregnancy (early normal pregnancy and abortion). The model had a Dice coefficient of 0.81, an accuracy of 0.66, and an AUC of 0.68. A comparative analysis with single-task models indicated that the proposed model outperformed them in both classification and segmentation tasks. CONCLUSION: Multitask learning simultaneously segments and classifies endometrium images, outperforming single-task models. This approach potentially accelerates ectopic pregnancy diagnosis from ultrasound scans before visualization of a gestational sac or serial β-hCG trends are established, improving early clinical detection and intervention.