A preprint reports higher benchmark segmentation scores for PROMISE-CNN than U-Net and generally higher scores for PROMISE-Txformer than UNETR across four medical-image datasets. They were ISIC-2017, Kvasir-Polyp, Kvasir-Instrument and CAMUS, covering skin lesions, colonoscopy polyps, endoscopic instruments and cardiac images.
The study proposes Prompt-Conditioned Channel Attention, or PCCA, as a hierarchical way to use prompts in an encoder-decoder segmentation model. It pools image and prompt features, maps their descriptors into a shared latent space and uses a gated excitation unit to generate prompt-aware channel weights.
The biggest gains came on some tests
Compared with U-Net, PROMISE-CNN reported Dice scores—an overlap measure—of 90.7% versus 82.2% on ISIC-2017 and 92.5% versus 86.0% on Kvasir-Polyp.
Its Dice scores were 97.2% versus 94.4% on Kvasir-Instrument and 91.3% versus 90.8% on CAMUS. The study also reported lower 95th-percentile Hausdorff distance, a boundary-error measure, and lower false-negative rates on the first three datasets.
PROMISE-Txformer generally scored higher than UNETR. Its Dice score rose from 82.5% to 89.0% on ISIC-2017 and from 88.6% to 90.0% on CAMUS, while the two Kvasir datasets showed higher Dice scores and lower HD95 in the reported comparisons.
The abstract reports intersection-over-union, or IoU, gains for PROMISE-CNN of 10.4% on ISIC-Lesion, 8.7% on Kvasir-Polyp, 0.8% on CAMUS-Cardiac and 3.4% on Kvasir-Instrument. For PROMISE-Txformer, the reported gains were 7.6%, 23%, 2.1% and 1.1% in the same dataset order. The supplied analysis says the displayed table differences need clarification against those reported percentages.
Tests of the model’s configuration
The strongest reported ablation metrics were listed for the version with PCCA at both the network bottleneck and decoder. It recorded a Dice score of 89.0% ± 8.2% and a false-negative rate of 10.6% ± 10.9% on ISIC-2017. On CAMUS, the corresponding figures were 90.0% ± 5.1% and 9.2% ± 5.8%. The comparisons were marked significant, although the exact statistical test was not reported.
In another test on ISIC-2017, the authors selected a channel-reduction ratio of 16 as the reported accuracy–efficiency compromise. That setting had a false-negative rate of 10.3%, an HD95 score of 12.8, 52.15 million parameters and 106.33 billion floating-point operations. Differences across the tested ratios were described as small.
Against the Segment Anything Model, or SAM, on ISIC-2017, PROMISE-CNN had a listed Dice score of 90.7%, compared with 85.1% for SAM. The table listed 118.15 and 61 for PROMISE-CNN, versus 2,991.32 and 312 for SAM, for floating-point operations and parameters respectively. This was a benchmark comparison, not deployment validation.
A closer look at the heart results
On the CAMUS cardiac benchmark, PROMISE-Txformer showed closer agreement with reference measurements than UNETR in the reported analysis. For end-diastolic volume, its mean bias—the average difference from the reference—was −6.7 millilitres, compared with −13.9 millilitres for UNETR. The reported 95% limits of agreement narrowed from −131.4 to 103.5 millilitres for UNETR to −95.3 to 81.9 millilitres for PROMISE-Txformer.
The correlation for ejection fraction was 80.2% for PROMISE-Txformer and 72.9% for UNETR. Mean bias in that measure was −2.08% for PROMISE-Txformer and −4.15% for UNETR. The agreement analysis does not establish prospective clinical interchangeability.
Why the findings remain preliminary
A prompt-robustness test used 300 ISIC-2017 test images. Bounding boxes from two observers, automatically generated boxes and synthetic shifts of plus or minus 20 pixels produced comparable performance, with pairwise Cohen’s d below 0.2. The test covered a limited subset and synthetic perturbations.
The comparisons were not uniformly conclusive: some Kvasir-Instrument differences were marked non-significant, and the table did not show a lower false-negative rate for PROMISE-Txformer there. The exact statistical tests were not reported for the ablation comparisons.
Paper data and sources
Original title: Prompt-Conditioned Channel Attention for Hierarchical Feature Modulation toward Anatomy-Agnostic Segmentation
Authors: Mosharof Hossain, Md Rabiul Islam, Limon Halder et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text