Work

PDC-Mask R-CNN: Pixel Difference Convolution for Enhanced Instance Segmentation of Multi-label Oral Lesions in Intraoral RGB Images

Year: 2026
Type: article
Abstract: Oral lesions in intraoral RGB images are small with indistinct boundaries, complicating segmentation.We propose PDC-Mask R-CNN, a multi-label instance segmentation framework that embeds Pixel Difference Convolution (PDC) into selected ResNet-50 residual blocks while leaving the prediction pipeline unchanged.PDC models interpixel intensity differences to strengthen boundaries against convolutional smoothing.We used 3,254 intraoral images (10,428 instances) across five categories: Calculus, Caries, Gingivitis, Ulcer, and Oral Cancer.Patient identifiers were unavailable, so we partitioned at the image level with near-duplicate screening, without claiming patient-level separation, and evaluated a screened test subset (303 images, 857 instances) under the COCO protocol.PDC-Mask R-CNN scored above the baseline on 19 of 22 metrics, including all eleven mask measures: mask AP@[0.50:0.95]rose from 16.695% to 19.317% (+2.622 pp, +15.7% relative), mask AP75 by 2.564 pp (+20.0%relative), and box AP by 1.771 pp (+9.7% relative).Three bounding-box metrics declined (large objects, Calculus, Gingivitis).Ablation attributes 0.276 of the 2.622 pp mask AP to the operator alone, raising AP75 but lowering AP50.Under the same protocol, PDC-Mask R-CNN exceeds the strongest of three evaluated one-stage segmenters, YOLO26l-seg, by 4.156 pp mask AP (+27.4% relative).All differences are descriptive single-run results; transformer-based and boundaryaware segmenters were not evaluated.
Language: English
Open Access status: hybrid
OpenAlex ID: W7235763609
FWCI:
Cites: 62
Cited by:
Related to: 10