Detection theory for accurate and non-invasive skin cancer diagnosis using dynamic thermal imaging

Abstract

Skin cancer is the most common cancer in the United States with over 3.5M annual cases. Presently, visual inspection by a dermatologist has good sensitivity (> 90%) but poor specificity (< 10%), especially for melanoma, which leads to a high number of unnecessary biopsies. Here we use dynamic thermal imaging (DTI) to demonstrate a rapid, accurate and non-invasive imaging system for detection of skin cancer. In DTI, the lesion is cooled down and the thermal recovery is recorded using infrared imaging. The thermal recovery curves of the suspected lesions are then utilized in the context of continuous-time detection theory in order to define an optimal statistical decision rule such that the sensitivity of the algorithm is guaranteed to be at a maximum for every prescribed false-alarm probability. The proposed methodology was tested in a pilot study including 140 human subjects demonstrating a sensitivity in excess of 99% for a prescribed specificity in excess of 99% for detection of skin cancer. To the best of our knowledge, this is the highest reported accuracy for any non-invasive skin cancer diagnosis method.

Keywords: (040.1490) Cameras; (040.1880) Detection; (040.3060) Infrared; (110.2970) Image detection systems; (170.1610) Clinical applications; (170.3660) Light propagation in tissues; (170.3880) Medical and biological imaging; (170.4580) Optical diagnostics for medicine; (330.1880) Detection.

Fig. 1

Tumor angiogenesis in cancer at different stages: (a) The tumor release growth factors that activate the growing cells generating blood vessel sprouts. (b) The blood vessels feed the tumor that growths thanks to cell proliferation. (c) The tumor becomes vascularized and it starts to metastasize through the blood stream (from webpage [33]).

Fig. 2

(a) Auto-covariance function for the null-hypothesis (H₀) estimated from patient data with known benign condition. (b) Auto-covariance function for the alternative-hypothesis (H₁) estimated from patient data with known malignant condition. In order to highlight their differences, (c) and (d) show the projection onto one of the left plane of the Auto-covariance function for the null-hypothesis and the alternative-hypothesis, respectively.

Fig. 3

False-alarm and detection probabilities parameterized by the threshold value, η, for different number of eigenfunctions used in the construction of the test-statistic (14)

Fig. 4

The theoretical receiver-operating characteristic (ROC) curve graphically shows the expected performance of the detector as we increase the number of eigenvalue-eigenfunction pairs. The larger the number of the pairs utilized to construct the test-statistic, the more statistical features utilized and the better the performance of the algorithm

Fig. 5

Block diagram of the detection stage of the proposed algorithm. The KL coefficients are computed by using the eigenfunctions of each hypothesis. These coefficients and the eigenvalues are used to compute the patient’s test-statistic, which is later compared with the optimum threshold to declare the malignancy

Fig. 6

ACFs for the case of vectorial random processes: (a) Autocorrelation function for the null-hypothesis (H₀) estimated from patient data with known benign condition. (b) Autocorrelation function for the alternative-hypothesis (H₁) estimated from patient data with known malignant condition.

Fig. 7

Acquisition hardware utilized to acquire the patient datasets. (a) Prototype and (b) Infrared imager and aquisition software

Fig. 8

Example of a patient dataset: (a) example of one square plastic marker used in the data acquisition step; (b) first frame of the infrared sequence, note that the visible and this frame are spatially aligned; and (c) the thermal recovery curves (TRCs) for the labeled pixels in (b).

Fig. 9

Comparison of the mean theoretical ROC curves over 200 permutations when 110 training are used to train the single-TRC algorithm (blue) and the dual-TRC algorithm (red). Comparison is made by using the mean AUC for different number of used eigenvalue-eigenfunction pairs, using 110 patients to train the algorithm.