HM_ADET: a hybrid model for automatic detection of eyelid tumors based on photographic images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate detection of eyelid tumors is crucial but challenging due to small, varied lesions and atypical early symptoms.
• Distinguishing benign from malignant eyelid tumors is difficult, especially for clinicians with limited experience, due to similar visual features.

Purpose of the Study:

• To develop a hybrid model (HM_ADET) for the automatic detection and classification of eyelid tumors.
• To improve the accuracy and efficiency of diagnosing benign versus malignant eyelid tumors.

Main Methods:

• Proposed a hybrid model, HM_ADET, integrating YOLOv7_CNFG for tumor localization and Vision Transformer (ViT) for classification.
• Utilized ConvNeXt module in YOLOv7_CNFG to preserve small tumor details and FReLU for multi-scale feature extraction.
• Employed GIoU loss for accurate bounding box prediction of irregular tumors and Multi-Head Attention in ViT for feature discrimination.

Main Results:

• The YOLOv7_CNFG component significantly improved Average Precision (AP) for eyelid tumor detection.
• YOLOv7_CNFG achieved AP of 0.893 (internal) and 0.765 (external), outperforming YOLOv7.
• The ViT classifier demonstrated high performance with Area Under the Curve (AUC) of 0.945 (internal) and 0.915 (external) for benign/malignant classification.

Conclusions:

• The HM_ADET model offers a promising strategy for automated eyelid tumor diagnosis.
• This automated approach has the potential to enhance patient outcomes and reduce healthcare expenses.