PnPD: Plug-and-Play Dual-Path Framework for Computer-Aided Detection with Decoupled Localization and Subtype Classification

Yung-Han Chen, Jian-Yu Jiang-Lin, Chen tsung hsing, Chang-Fu Kuo, Hung-Yu Wu, Wen-Huang Cheng, Juinn-Dar Huang

Keywords: Colonoscopy, object detection, semantic segmentation

Abstract: Colonoscopy-based computer-aided detection (CADe) requires reliable polyp localization alongside clinically meaningful subtype classification. However, modern single-stage detectors optimize both tasks on a shared representation, creating a fundamental feature conflict: localization relies on boundary cues, while subtyping demands intra-lesion fine-grained textural details. To resolve this inherent limitation, we propose PnPD, a Plug-and-Play Dual-path framework that decouples localization and classification into independent, parallel streams. This modular design allows for the flexible integration of specialized models, utilizing lightweight object detectors exclusively for bounding box regression while reformulating subtype classification as a segmentation task. By replacing the standard classification head with a segmentation module, our approach explicitly forces the model to capture dense, pixel-level mucosal patterns. Extensive experiments demonstrate that our framework effectively resolves the feature misalignment observed in unified models. Remarkably, lightweight PnPD configurations match or surpass the detection F1-scores of large and x-large YOLO baselines while reducing parameters and MACs by over 80%, simultaneously improving average subtype precision and recall by more than 9% and 7%. This offers a scalable solution tailored for efficient deployment in real-time clinical applications.

Primary Subject Area: Detection and Diagnosis

Secondary Subject Area: Application: Endoscopy

Registration Requirement: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 54