Abstract
Early Esophageal Cancer (EEC) segmentation is a critical yet challenging task in clinical endoscopy due to the irregular boundaries of lesions, extreme data scarcity, and stringent patient privacy regulations. We present YOHO ("You-Only-Have-One"), a novel deep learning framework that shifts the paradigm from general-purpose models to patient-specific optimization. By training a dedicated neural network for a single patient image using a "cut-and-paste" data augmentation strategy, YOHO achieves a mean Dice score of 0.888 on EEC datasets, significantly outperforming state-of-the-art models while inherently preserving data privacy.
Introduction
Early Esophageal Cancer (EEC) is notoriously difficult to diagnose and segment. Unlike colorectal polyps, which often possess distinct morphologies, EEC lesions are "shape-shifters" with complex, irregular boundaries that frequently blend into surrounding healthy tissue. Standard deep learning approaches encounter three primary obstacles in this domain: data scarcity, the generalization gap, and privacy concerns. High-quality annotated EEC datasets are rare, and models trained on one patient often fail to generalize to others due to variations in lighting, angle, and tissue appearance. Furthermore, sharing medical data for centralized training is often restricted by legal and ethical frameworks.
To address these issues, the YOHO framework adopts a "one-image-one-network" philosophy. Instead of building a universal model, YOHO optimizes a specific model for a single high-quality frame, effectively turning the traditional problem of "overfitting" into a clinical advantage.
Methodology: The YOHO Framework
The YOHO workflow consists of three integrated stages designed to maximize the information extracted from a single clinical image: interactive rendering, edge-enhanced training, and intentional overfitting.
Interactive Rendering
The process begins with a clinician providing a rough, "quick and dirty" sketch of the lesion area. The system then samples small circular patches from this region. These patches are transformed into simple geometric shapes (circles and triangles) and randomly pasted onto the healthy regions of the same image. This process generates a synthetic training set of approximately 1,600 images, teaching the network to recognize the specific texture and color characteristics of that patient's lesion against their healthy tissue.
Edge-Enhanced UNet (eUNet)
The core architecture is a modified UNet, termed eUNet, which incorporates a dedicated Edge Detector branch and a Consistency Loss function. This design forces the network to prioritize boundary detection, which is essential for the ambiguous edges typical of esophageal lesions.
Intentional Overfitting
By restricting the model's scope to a single image, YOHO can be pushed to over-fit to that specific patient's unique tissue characteristics. This results in extreme precision that general models, which must account for inter-patient variability, cannot match.
Experimental Results
The YOHO framework was evaluated on the EEC-2022 dataset (1,236 images) and standard polyp datasets (CVC-612 and Kvasir-SEG).
Performance on EEC
YOHO achieved a mean Dice score of 0.888 on the EEC-2022 dataset. This represents a substantial leap over existing state-of-the-art (SOTA) models. For instance, Polyp-PVT and Swin-UNet, which are highly effective in other medical imaging tasks, struggled to exceed a 0.74 Dice score on the same data. Qualitative comparisons demonstrate that YOHO produces much smoother and more anatomically accurate boundaries.
Comparison with Foundation Models
The researchers also compared YOHO against Meta’s Segment Anything Model (SAM). Despite SAM's massive pre-training, YOHO significantly outperformed it in Dice scores, proving that domain-specific, localized optimization is superior to general-purpose foundation models for specialized medical tasks.
Discussion and Conclusion
YOHO introduces a "Privacy by Design" architecture; since no external patient data is required to train a specific model, the risk of data leakage is eliminated. While the method requires a human-in-the-loop for the initial rough sketch and approximately 4.5 minutes of training time on an RTX 2080 Ti, these requirements are manageable within post-procedure clinical workflows.
Future work will focus on reducing training latency to enable real-time intra-operative use and extending the YOHO philosophy to other modalities like CT and MRI. By embracing the unique characteristics of individual images, YOHO provides a robust solution to the long-standing challenges of medical image segmentation.
Summary updated