This article summarizes work our team published in Computers in Biology and Medicine, co-authored with collaborators at the Universitat de Barcelona, University College London, and the Universidad de Oviedo. It is a good window into how we think about medical imaging AI: the hard part is rarely the model — it is the messiness of real clinical data.
The problem: not all images are equal
Ulcerative colitis is a chronic disease that needs regular monitoring, and today that usually means invasive colonoscopy. Transabdominal ultrasound is a promising non-invasive alternative — it can image the colon wall without a procedure — but interpreting it takes real expertise, which is exactly what makes it a candidate for automation.
The catch is quality. Ultrasound image quality swings dramatically with patient anatomy, operator skill, and machine settings. Speckle noise and artifacts obscure the very structures you are trying to trace, and the colon wall is a small target that looks a lot like the tissue around it. Standard segmentation models are trained as if every image were clean, so they degrade badly on the low-quality scans that are common in practice — and those are often the cases that matter most.
The idea: make the model quality-aware
Rather than hiding from image quality, we made the model predict it. Our architecture is a U-Net with a ConvNeXt encoder and a second, parallel branch that estimates the quality of each image while it segments. That quality signal does real work: it weights the loss during training, so the model leans on reliable, well-annotated images and discounts uncertain labels on the noisiest ones.
This matters because on a bad image, even human experts disagree about where the colon wall is. Forcing the model to treat every annotation as ground truth teaches it to imitate that disagreement. By explicitly modeling quality, the network learns robust features from the images it can trust and stops overfitting to noise it cannot.
The results: past expert-level agreement
We evaluated on the public C-TRUS dataset and compared against state-of-the-art approaches, breaking results out by image quality. Our method reached Dice scores of 0.78, 0.70, and 0.60 for high-, medium-, and low-quality images — with the biggest gains exactly where they are hardest to get, on the low-quality scans.
The number we find most telling: the original dataset reported an average inter-observer agreement among clinical experts of 0.61 Dice. Our framework's overall performance exceeds that — the model agrees with the reference more consistently than the experts agree with each other. That is the threshold where automated segmentation starts to become clinically useful rather than merely academically interesting.
Why this reflects how GI works
Two habits from this project show up in every medical-imaging engagement we take on. First, we treat data quality as a first-class variable, not an afterthought — because in the real world it is the thing that breaks models after they leave the lab. Second, we benchmark against the right baseline: not a demo number, but how well human experts actually agree, which is the bar a clinical tool has to clear.
The work is honest about what is left, too. Broader multi-institutional validation, objective quality metrics, and clinical studies on patient impact all remain ahead. That is the difference between a published method and a deployed medical product — and being clear about which one you have is part of doing this responsibly.