Purpose: Minimizing the false negative rate (FNR) is critical because it directly influences the early diagnosis of melanoma-a key determinant in improving patient survival rates.This study aimed to reduce melanoma misdiagnosis by analyzing the FNR minimization in a convolutional neural network (CNN)-based melanoma classification model using hyperspectral datasets.Methods: We obtained 32 hyperspectral imaging (HSI) samples from 20 patients with biopsyconfirmed diagnoses.Dermatologists manually delineated regions of interest for normal skin, melanoma, and basal cell carcinoma.A total of 254,433 pixels were extracted from the 32 HSI samples, and a one-dimensional CNN was trained using these pixel-level spectral data.Additionally, 1,800 dermoscopy images were retrieved from the International Skin Imaging Collaboration 2019 dataset via the Kaggle platform.EfficientNet B7 was trained using dermoscopy images, and separate CNN models were developed for each dataset.Model performance was evaluated for sensitivity, specificity, area under the curve (AUC), and FNR.Furthermore, the FNR of the CNN-based model using hyperspectral data was indirectly compared with values reported in previous studies to assess the reliability of our findings. Results:The HSI-based model achieved a melanoma FNR of 0.0284, with a sensitivity of 0.9716, specificity of 0.9883, and AUC of 0.9799.In comparison, the dermoscopy-based model yielded a melanoma FNR of 0.1667, sensitivity of 0.8333, specificity of 0.8760, and AUC of 0.9090. Conclusions:The CNN-based melanoma classification model using hyperspectral data demonstrated a substantial reduction in melanoma FNR.These findings support the feasibility of integrating HSI with CNNs to improve melanoma detection accuracy.
Heo et al. (Thu,) studied this question.