A Two-Stage Cascaded Regression Framework for Automatic Facial Acupoint Localization in Infrared Thermal Images

Area of Science:

• Medical Imaging
• Computer Vision
• Traditional Chinese Medicine (TCM)

Background:

• Infrared thermal imaging provides objective physiological data for TCM.
• Automated acupoint localization faces challenges with low texture and extreme pose variations in thermal images.

Purpose of the Study:

• To develop an automated acupoint localization framework for thermal facial images.
• To address limitations of low texture and pose variations in existing methods.

Main Methods:

• Constructed a multi-pose thermal facial dataset using digitized bone-proportional measurements and TCM anatomical rules.
• Proposed T2FAL, a two-stage cascaded regression framework with a Thermal-Aware Face Detector (Stage 1) and pose-specific regressors (Stage 2).
• Stage 1 utilized an improved YOLOv12m with ICAN_C2f, MixNeck, and TAF-IoU. Stage 2 incorporated Gated Feature-Conditioned Cascade Refinement (FCCR) and a Selective GeoDeriv module.

Main Results:

• Stage 1 achieved 87.3 frames per second on an NVIDIA RTX 4090.
• Stage 2 frontal-view configuration yielded mAP@50-95 of 73.19% and mean pixel error of 1.986 pixels.
• Stage 2 profile-view configuration yielded mAP@50-95 of 86.92% and mean pixel error of 3.109 pixels.

Conclusions:

• The T2FAL framework demonstrates the feasibility of automated reference-coordinate localization on thermal facial datasets.
• Further validation with independent expert annotations, external test sets, and clinical data is necessary for diagnostic use.