Actinic keratosis staging in multimodal image data.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Actinic Keratosis (AK) is a common sun-induced skin condition with potential for cancer progression.
• Early detection and staging of AK are critical for timely intervention.
• Characteristic dermatoscopic and structural changes occur during AK progression.

Purpose of the Study:

• To develop a multimodal, machine-learning framework for automated AK staging.
• To integrate dermatoscopic and high-frequency ultrasound (HFUS) data for enhanced lesion identification.
• To identify early-stage AK lesions accurately.

Main Methods:

• A dataset of 222 paired dermatoscopic and HFUS images was analyzed.
• Features were extracted from both imaging modalities, including texture and structural data.
• Various machine learning models (Logistic Regression, RF, SVM, MLP) and a TwinCNN were evaluated for classification.

Main Results:

• Multimodal models integrating dermatoscopic and HFUS data outperformed single-modality approaches.
• High accuracy (over 80%) was achieved for classifying healthy skin versus AK stages (AK1-AK3).
• Early AK detection against healthy and advanced AK reached nearly 90% accuracy using multimodal features.

Conclusions:

• A multimodal imaging framework combining dermatoscopy and HFUS provides accurate AK classification.
• This approach surpasses the performance of single-modality imaging techniques.
• Future research should focus on larger datasets and advanced fusion architectures.