MATERIALS AND METHODS

Medical applications of thermography: Thermal imaging finds applications in breast oncology, integrative medicine, plastic surgery, dentistry, orthopedics, acupuncture, occupational medicine, pain management, vascular medicine, cardiology and veterinary medicine. It has been identified to be the best technique for preliminary breast examinations. Clinical applications employ static and dynamic thermography methods for capturing the thermal profile of the subjects. The provision for both qualitative and quantitative assessment in thermography helps in the estimation of progression of the disease in a systematic manner. The results of clinical assessment by facial thermography are discussed in detail by Gratt et al. (1996). From the results it is observed that, hot, cold and normal facial thermograms give a clear indication of sinusitis, peripheral nerve mediated pain and psychogenic facial pain , respectively. Thermography is used in the diagnosis and treatment of chronic orofacial pain patients, Temporo Mandibular Joint (TMJ) disorders etc. (Sudhakar et al., 2011). The study concludes that thermography aids in the assessment and staging of various dysfunctions of the head and neck region. Medical thermography is followed as an adjunctive procedure along with other diagnostic procedures to help the clinician in making a thorough examination. Lately, the significance of selection of acupoints aided by facial thermography for treating facial paralysis has been reported by Nan and Hou (2012). The studies with subjects indicate that, acupuncture treatment provided at selected acupoints aided by thermography resulted in fewer sessions of short treatment durations.

Medical image authentication with watermarking: The significance of medical image watermarking for integrity and authenticity has been elaborately discussed in a study by Coatrieux et al. (2006). Zhou et al. (2001) presented a scheme for watermarking digital mammography images with a message digest generated with MD5 algorithm for enforcing integrity. Lossless or reversible (Tian, 2002) watermarking schemes have invited attention owing to the need for high quality images with no or less degradation for medical diagnosis, military investigation and forensic analysis etc. As medical images contain sensitive data which an expert cannot afford to lose, reversible watermarking schemes are applied to recover the medical images after the extraction of the watermark. The strength of a watermarking system can be improved by incorporating the paradigm of blind (Bartolini et al., 1999) watermarking which refers to the watermark extraction without the need for the original unmarked image during the extraction procedure. It prevents the unauthorized distribution and copying of the original images. A review of reversible watermarking techniques, their classification and applications have been elaborately discussed in an article by Caldelli et al. (2010). This article summarizes various reversible watermarking techniques which are based on simple integer arithmetic operations.

Thermal visual image fusion: Often the thermal and visual images are compared by the clinician to localize the pathology based on the thermal patterns. This procedure is followed for medial diagnosis as well as prognosis. It requires the clinician to perform a mapping between these images. However, a composite image of these modalities will help the clinician to cross reference the thermal profile of the subject with that of the surface anatomy. A method for overlay has been presented by Schaefer et al. (2006) based on skin detection and image registration. From medical literatures, once can identity the significance of fusion for treating subjects with Frey’s Syndrome (FS). This is characterized by excessive sweating of the forehead, upper lip, peri-oral region or sternum subsequent to intake of food. Isogai and Kamiishi (1997) have proposed thermography as a diagnostic procedure for FS. Localization of the foci of FS in the facial region is essential to administer the Botulinum Toxin (Botox) or performing parotid surgery. Kragstrup et al. (2011) presented an elaborate literature review on FS and a case report of a subject with FS. The authors conclude that facial thermograms and starch iodine tests supplement the clinical examination of the subject.

Proposed system: The proposed system is the extension of the work on watermarking a medical image with the facial image of the subject for improved telediagnosis and tamper detection (Velumani and Seenivasagam, 2010). The previous work is extended to demonstrate the fusion of the host and watermark images for enhanced diagnosis. The proposed systems is termed as semiblind as the watermark extraction algorithm needs a key for watermark extraction and reconstruction of the original host image. The algorithms for watermark embedding and extraction and superimposition are given below.

Watermark embedding and extraction: The algorithms are based on the application of a Triangular Number Generator (TNG) function. It is a function as in (1) which uniquely encodes a pair of positive integers (a, b) into a unique number TR. The number TR can be factored back to a and b exactly by applying the extraction procedure. The proposed system exploits this reversible and blind nature of the TNG in the watermark embedding and extraction algorithms. The properties of TNG are discussed in detail in an article by Gupta (2002).

(1)

Watermark embedding is performed by combining n number of LSB planes of the thermal and visual images by Eq. 1. As the combination leads to an integer whose magnitude cannot be expressed in n bits, modulo 2ⁿ operation is performed on the resultant integer. While the n bit reminder is appended with 8-n bit planes, the quotient is preserved as a key. Similarly, on watermark extraction, the combination is reconstructed by combining the key and the n LSB planes of the watermarked image. The watermark embedding and extraction algorithms are given in Algorithm 1 and 2, respectively assuming n = 4.

Algorithm 1 watermark embedding:

Algorithm 2 watermark extraction:

The above algorithms are illustrated for with Fig. 1 and 2, respectively. The facial profile of a subject with FS who has taken a starch iodine test is taken as the visual image V and the thermogram of the same area is taken as the host image T.

Fig. 1:	Watermark embedding

Fig. 2:	Watermark extraction

Visual and thermal image overlay: The superimposition of the visual and thermal images is performed by computing the weighted summation of the intensity values of the thermal and visual image pairs. As the purpose of overlay is to localize the area of pathology in the human anatomy, the pixel intensity values of the thermal and visual images are scaled down by a factor α and β, respectively. The resultant value is added with the pixel intensity value in the same spatial orientation of the visual image to get the composite intensity value. The overlay generation is given in Algorithm 3 as below.

Algorithm 3 thermal and visual image overlay:

RESULTS AND DISCUSSION

The proposed algorithms have been implemented in MATLAB with the visual and thermal images of healthy subjects and subjects with pathologies. For healthy subjects the facial images have been chosen as watermarks. For thermograms with pathologies, the digital images of the surface anatomy have been chosen as watermarks. All the images are of the size 512x512x3. The results are illustrated with performance metrics for healthy subjects and subjects with pathologies in Table 1 and 2, respectively. The Peak Signal to Noise Ratio (PSNR), a measure of the image perceptibility is evaluated for the watermarked and reconstructed images and the watermarks. The composite images are given in the last row of the tables. Invariably the values of α and β are assumed to be .5 and .7, respectively for all images to generate the overlays.

The thermal and visual images in Table 1, produced by Mikos Biotek in Virginia, USA are taken from the Science Photo Library database (Biotek, 2012). The visual and thermal images in Table 2 are taken from the gallery of IR images (Mercer, 2011).

Table 1:	Test images and experimental results for healthy subjects

Table 2:	Test images and experimental results for pathologies

Table 3:	Comparison with earlier approaches
The above table compares the proposed scheme with the earlier schemes reviewed in the article by Caldelli et al. (2010) with respect to the embedding capacity and PSNR values

The experiments have been implemented for the following pathologies:

Exclusively, the visual image illustrating the Frey’s syndrome is one which shows black patches in the parotid area as a result of Minor’s Starch Iodine test.

It is seen from Table 1 that the PSNR values of the watermarked images are closer to 38 dB for facial thermograms. The watermarked thermogram images in Table 2 seem to have PSNR values around 36 dB. Though the PSNR values are not significantly high, 36 dB is an acceptable value for medical images. PSNR is also attributed by the size of the watermark. In this experiment both the visual and thermal images are resized to the same dimension before watermarking for pixel by pixel alignment for the generation of overlay. The PSNR values for the watermarked images will significantly improve on reducing the dimension of the watermark. However, it will necessitate another resize for overlay. It is clearly understood from the experimental results that the proposed system provides perfect reconstruction of the cover image and authentication of the thermogram and patient identification without ambiguity. It is evident from the PSNR values which are Infinity for both the recovered cover images and the watermarks for both healthy subjects and pathologies.

The proposed system based on spatial watermarking is compared with similar systems discussed in the review article by Caldelli et al. (2010). The performance of the proposed system is compared with the other schemes in respect of the PSNR values of the watermarked images. As arbitrary images have been taken in different studies, the comparison is based on the size of the cover images and the watermark. The details of the comparisons are tabulated in Table 3. Only the best PSNR values realized by the other schemes are taken for comparison.

From the Table 3, it is evident that the proposed system provides a theoretical embedding capacity of 4 bpp (bits per pixel) with an average PSNR of 37 dB. In contrast to the above schemes, the semi-blind nature of the proposed system requires the secured storage and exchange of the key which is of the size 512x512x3. Though, it is an overhead, it is worth in accomplishing the watermark extraction and cover image recovery in an intact manner. Further, as the key is a combination of the cover and the watermark images, guessing the key by an attacker is highly unrealistic and it provides a great deal of security.

In addition to the above, the method for superimposition presented in the study seems to be a promising one for generating the overlays. The proposed scheme for overlay generation is contrary to the scheme proposed by Schaefer et al. (2006), in which the final overlay image is weighted by 20 and 80% of the thermal and visual images, respectively. The pixel by pixel superimposition proposed in this study guarantees the perfect geometric alignment of the visual and thermal patterns. Particularly, in the case with Frey’s syndrome, the overlay generated by the proposed approach seems to exactly localize the region for injection of Botox than the conventional starch iodine test. Hence, the preciseness of the watermarking system complements the overlay generation.

CONCLUSION

The proposed semiblind reversible watermarking scheme for authenticating the thermograms contributes to generate a perfect overlay of the thermal and visual images. The proposed system is novel when compared to the earlier reversible watermarking systems in the approach selected for realizing reversibility and blind extraction. From the diverse nature of the thermal images chosen, it found that the proposed system can be extended to all thermogram based diagnosis and prognosis. The system can be further enhanced by eliminating the need for the key. However, the semiblind nature of the system protects the images from unauthorized access and also indicates incidental or accidental tampering. The efficacy of the proposed system can be tested further by experimenting with medical images related to fine scale dermatitis, melanoma and other syndromes which require exact localization for incision in surgical procedures.