Clinical study of noninvasive in vivo melanoma and nonmelanoma skin cancers using multimodal spectral diagnosis

Abstract

The goal of this study was to determine the diagnostic capability of a multimodal spectral diagnosis (SD) for in vivo noninvasive disease diagnosis of melanoma and nonmelanoma skin cancers. We acquired reflectance, fluorescence, and Raman spectra from 137 lesions in 76 patients using custom-built optical fiber-based clinical systems. Biopsies of lesions were classified using standard histopathology as malignant melanoma (MM), nonmelanoma pigmented lesion (PL), basal cell carcinoma (BCC), actinic keratosis (AK), and squamous cell carcinoma (SCC). Spectral data were analyzed using principal component analysis. Using multiple diagnostically relevant principal components, we built leave-one-out logistic regression classifiers. Classification results were compared with histopathology of the lesion. Sensitivity/specificity for classifying MM versus PL (12 versus 17 lesions) was 100%/100%, for SCC and BCC versus AK (57 versus 14 lesions) was 95%/71%, and for AK and SCC and BCC versus normal skin (71 versus 71 lesions) was 90%/85%. The best classification for nonmelanoma skin cancers required multiple modalities; however, the best melanoma classification occurred with Raman spectroscopy alone. The high diagnostic accuracy for classifying both melanoma and nonmelanoma skin cancer lesions demonstrates the potential for SD as a clinical diagnostic device.

Fig. 1

Spectral diagnosis (SD) system in a clinical setting. It consists of two independent systems, each with a customized fiber optic probe. Details of the system are available in Sec. 2.

Fig. 2

Mean spectra of melanoma (MM) nonmelanoma pigmented lesions (PL), and normal skin. One of the melanoma lesions is an amelanotic melanoma (AM): (a) RS, (b) DOS, and (c) LIFS.

Fig. 3

Mean spectra by pathology for nonmelanoma skin cancer (NMSC; BCC, SCC, and AK) compared with normal skin: (a) RS, (b) DOS, and (c) LIFS.

Fig. 4

Receiver operating characteristic curves for all classifiers, with corresponding area under the curve (AUC) shown in legend. The sensitivity and specificity for each classifier are marked. We use per lesion analysis, described in Sec. 2.

Fig. 5

Effect of standardization on DOS (a, b) and LIFS data (c, d): DOS prestandardization (a) and poststandardization (b), and LIFS prestandardization (c) and poststandardization (d).

Fig. 6

RS standardization to AUC of amide I peak (1642 to $1660{\mathrm{cm}}^{-1}$). (a) RS prestandardization and (b) RS poststandardization.

Fig. 7

PC scores (D1 and D2) for classifying BCC versus normal (N) using per measurement analysis (a) versus per lesion analysis (b). For better visualization, this plot zooms at the region around the decision line. Legends: TN = true negative (normal skin measurements on the negative side of the decision line), PLFP = per lesion false positive (normal skin measurements with at least one measurement on the positive side of the decision line), TP = true positive (BCC measurements on the positive side of the measurements), PLFN = per lesion false negative (all measurements from the same BCC lesion located on the negative side of the decision line), and PLP = per lesion positive (BCC measurements that have a corresponding lesion measurement on the positive side of the decision line).