Multiharmonic imaging-based automated recognition of cutaneous T-cell lymphoma

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Cutaneous T-cell lymphomas (CTCL), particularly mycosis fungoides (MF), present diagnostic challenges in early stages due to nonspecific clinical and histopathologic findings.
• Delayed diagnosis of MF can lead to significant delays in treatment, impacting patient outcomes.
• Novel diagnostic methods are crucial for improving the diagnostic and therapeutic strategies for CTCL.

Purpose of the Study:

• To evaluate Nonlinear Optical Microscopy (NLOM) for imaging hematoxylin & eosin (H&E) stained skin samples.
• To develop and apply an artificial intelligence (AI) model for detecting atypical epidermotropism and dermal cells in MF skin samples using NLOM.
• To assess the potential of NLOM for in vivo skin imaging of fresh, unstained biopsy samples.

Main Methods:

• Analysis of H&E-stained skin biopsy samples from MF lesions using both brightfield microscopy and NLOM.
• Training a convolutional neural network (CNN) with expert-labeled images to identify skin lymphocytes.
• Application of the AI model to independent datasets from both imaging modalities and fresh, unstained samples.

Main Results:

• NLOM successfully imaged epidermal and dermal structures in H&E-stained MF tissue with sub-cellular resolution.
• The AI model accurately detected lymphocyte epidermotropism and dermal infiltration in analyzed images.
• NLOM demonstrated capability for imaging fresh, unstained biopsies up to 400 µm deep.

Conclusions:

• NLOM coupled with AI effectively detects key features of MF in H&E-stained skin tissue, including lymphocyte epidermotropism and dermal infiltration.
• This combined approach offers a valuable tool for dermatologists to enhance MF-CTCL diagnosis and prognosis.
• The findings suggest potential for more timely and precise therapeutic interventions in MF-CTCL management.