Healthcare & Bio AI
New attention-guided model outlined lesions in three medical image datasets more accurately than comparable methods
In a 2025 paper in Knowledge-Based Systems, researchers tested a new biomedical image segmentation model called DCSSGA-UNet on three medical image datasets: CVC-ClinicDB, CVC-ColonDB and Kvasir-SEG. Segmentation is the task of marking the outline of a region, such as a lesion, in an image. The dice coefficient and intersection-over-union both measure how closely the predicted outline overlaps with a reference outline, and higher scores mean closer overlap. The model reported mean dice scores of 98.85%, 95.71% and 96.10%, and mean intersection-over-union scores of 95.67%, 92.39% and 93.97% on the three datasets. The authors report that it outperformed the comparable methods they tested.
What kind of model is it?
DCSSGA-UNet uses DenseNet201, a type of convolutional neural network designed to process images, as its base encoder. That is the part of the network that reads the image and extracts features. It adds two attention blocks, called channel spatial attention and semantic guidance attention, which weigh useful features more heavily and reduce redundant information. The authors say this helps the model handle lesions that vary in size and shape.
How well did it perform?
The paper reports three pairs of results. On CVC-ClinicDB, the model achieved a mean dice score of 98.85% and a mean intersection-over-union score of 95.67%. On CVC-ColonDB, it achieved 95.71% and 92.39%. On Kvasir-SEG, it achieved 96.10% and 93.97%. The authors report that DCSSGA-UNet outperformed the comparable methods they tested.
What might this mean for patients?
The authors describe the model as potentially useful for clinical applications, especially for outlining lesions and assisting with diagnosis and treatment of diseases such as colorectal cancer. The paper reports benchmark image results, not a clinical trial, so this is an early technical result.
What else should I know?
What is biomedical image segmentation?
Biomedical image segmentation is marking the outline of a region, such as a lesion, in a medical image. This study tested a model that performs that task.
Did the researchers test on patients?
No. The abstract describes benchmark tests on three medical image datasets, not clinical tests with patients.
Where was the study published?
The paper appeared in Knowledge-Based Systems in 2025.
Limits and what's next
The abstract describes a benchmark study on three medical image datasets: CVC-ClinicDB, CVC-ColonDB and Kvasir-SEG. The abstract does not describe any independent replication or clinical validation. The authors present the model as a technical result that may assist clinical work. More evaluation would be needed to show how it performs in real clinical settings.
Source: Tahir Hussain et al., “DCSSGA-UNet: Biomedical image segmentation with DenseNet channel spatial and Semantic Guidance Attention,” Knowledge-Based Systems, 2025.