Cross-Organ Domain Adaptive Neural Network for Pancreatic Endoscopic Ultrasound Image Segmentation

TL;DR

This work tackles the data scarcity and cross-organ domain gap in pancreatic endoscopic ultrasound image segmentation by proposing COTS-Nets, a dual-network framework that transfers cross-organ tumor knowledge. The universal network is guided by a boundary loss and an EMA-based consistency loss to align lesion representations across organs, while an auxiliary network with a Hybrid Adaptive Attention Module extracts domain-invariant, multi-scale features and distills them into the universal network. The method is validated on cross-organ adaptations using MMOTU_2d and Thyroid_tn3k as sources and the PCEUS pancreatic dataset as the target, showing consistent improvements in Dice, IoU, ASD, and HD95 over state-of-the-art methods; ablations confirm the contributions of boundary learning, consistency guidance, and cross-organ HAAM features. The Pancreatic Cancer Endoscopic Ultrasound (PCEUS) dataset of 501 images is introduced to facilitate future research. Overall, COTS-Nets demonstrates that cross-organ tumor homogeneity can be leveraged to enhance segmentation in data-scarce EUS settings with practical implications for improved pancreatic cancer diagnosis.

Abstract

Accurate segmentation of lesions in pancreatic endoscopic ultrasound (EUS) images is crucial for effective diagnosis and treatment. However, the collection of enough crisp EUS images for effective diagnosis is arduous. Recently, domain adaptation (DA) has been employed to address these challenges by leveraging related knowledge from other domains. Most DA methods only focus on multi-view representations of the same organ, which makes it still tough to clearly depict the tumor lesion area with limited semantic information. Although transferring homogeneous similarity from different organs could benefit the issue, there is a lack of relevant work due to the enormous domain gap between them. To address these challenges, we propose the Cross-Organ Tumor Segmentation Networks (COTS-Nets), consisting of a universal network and an auxiliary network. The universal network utilizes boundary loss to learn common boundary information of different tumors, enabling accurate delineation of tumors in EUS despite limited and low-quality data. Simultaneously, we incorporate consistency loss in the universal network to align the prediction of pancreatic EUS with tumor boundaries from other organs to mitigate the domain gap. To further reduce the cross-organ domain gap, the auxiliary network integrates multi-scale features from different organs, aiding the universal network in acquiring domain-invariant knowledge. Systematic experiments demonstrate that COTS-Nets significantly improves the accuracy of pancreatic cancer diagnosis. Additionally, we developed the Pancreatic Cancer Endoscopic Ultrasound (PCEUS) dataset, comprising 501 pathologically confirmed pancreatic EUS images, to facilitate model development.

Figures (7)

• Figure 1: The endoscopic ultrasound (EUS) procedure for detecting pancreatic masses involves using the EUS device to examine the pancreatic lesion. The area depicted by the green curve in the figure is the identified mass.
• Figure 2: The training process for radiologists is not limited to a specific organ. Initially, they gain meta knowledge by learning to identify diagnostic lesions across various regions. After completing their studying, radiologists typically focus on diagnosing lesions in a specific organ.
• Figure 3: The schematic illustration of our proposed Cross-Organ Tumor Segmentation Networks (COTS-Nets), which is trained on cross-organ data using encoder-decoder with auxiliary network. The universal network comprises both encoder and decoder components to mitigate cross-organ heterogeneity. In each training iteration, the framework processes multiple batches of images, with each batch corresponding to a different domain. The universal network is trained with the supervision of ground-truth labels and the transferred knowledge from the auxiliary branches.
• Figure 4: The components of auxiliary network branch.
• Figure 5: T-SNE distribution visualization. The MMOTU_2d data is represented in green, the Thyroid_tn3k data in blue, and our self-constructed PCEUS dataset in orange.