A New Approach to Skin Cancer Diagnosis: Research on Photoacoustic Imaging with Level Set Segmentation Algorithm

In recent years, with global population aging and environmental changes, the incidence rate of skin cancer has been climbing annually. As a significant public health issue, skin cancer primarily manifests in non-melanoma types like squamous cell carcinoma (SCC) and basal cell carcinoma (BCC). Among these, basal cell carcinoma is the most common type. Statistics indicate that approximately 4.3 million new cases of BCC are reported annually in the United States. Although BCC has a lower mortality rate, it poses tremendous challenges to patients’ quality of life and healthcare resources.

Clinical diagnosis and treatment of BCC still encounter many challenges. Traditional methods for assessing tumor boundaries, such as histopathology, are accurate but require invasive procedures like biopsies and consume significant time. In addition, advanced non-invasive imaging techniques such as Optical Coherence Tomography (OCT) and Reflectance Confocal Microscopy (RCM) offer high resolution but are limited in tissue penetration depth (RCM: ~200–300 μm, OCT: 1–2 mm), making them less effective for comprehensively mapping deeper tumor boundaries. These limitations hinder broader application in clinical management of various BCC subtypes.

Against this backdrop, an innovative study led by Xiuting Li, Valerie Xinhui Teo, and others proposed a new approach combining Multispectral Optoacoustic Tomography (MSOT) with a Level Set Segmentation Algorithm to accurately and non-invasively map pigmented BCC in Asian skin. This research, published in the European Journal of Nuclear Medicine and Molecular Imaging, was a collaborative effort between the Skin Research Labs at the Agency for Science, Technology and Research (A*STAR) in Singapore and the National Skin Centre of Singapore.

Background and Methodology

This study aims to address the challenges of dynamic monitoring of tumor boundaries and overcome the inadequate penetration depth and contrast of existing optical imaging methods. By developing and integrating an automated level set image segmentation approach in MSOT, the researchers aim to provide an imaging solution capable of precise measurements of tumor width, depth, and volume, laying the foundation for preoperative mapping and surgical planning.

Clinical Study Design

The study was approved by the Domain Specific Review Board (DSRB) of the National Health Group and the A*STAR IRB (Clinical Study Reference Numbers: ²⁰²⁰⁄₀₀₁₁₅ and ²⁰²²⁄₀₀₃₄₇). A total of 65 patients, aged 21–90 years, diagnosed with non-melanoma skin cancer and scheduled for excision or Mohs Micrographic Surgery (MMS), were recruited. After histopathological confirmation, 30 patients with pigmented BCC became the final research cohort. All participants were classified as Fitzpatrick skin types III–IV.

The MSOT imaging system used was the MSOT Acuity device by iThera Medical GmbH, equipped with a handheld 3D probe. The system supports a wavelength range of 680–980 nm, with a maximum penetration depth of 10 mm and a spatial resolution of 80 μm.

Imaging and Data Collection Workflow

1. Imaging Preparation: Written informed consent was acquired from patients, followed by the collection of clinical and dermoscopic images.
2. MSOT Imaging: The handheld probe was placed at the target site, and photoacoustic signals were collected and spectrally unmixed (oxyhemoglobin, deoxyhemoglobin, and melanin signals) within the wavelength range of 680–920 nm.
3. Post-Imaging Marking and Sample Processing: After imaging, a surgical pen marked tumor locations for accurate alignment of biopsy samples.
4. Data Processing and Analysis: The acquired photoacoustic data were reconstructed into 100×100×100-pixel 3D images for subsequent analysis.

Image Processing and Algorithm Development

Preprocessing

The MSOT-acquired images were first preprocessed using Maximum Intensity Projection to project 3D images onto (x-y, x-z, y-z) planes, and a Median Filter was applied to reduce random noise.

Level Set Segmentation Algorithm

The level set method applies convex relaxation from the continuous Potts model to refine and annotate BCC boundaries accurately. Key steps include: 1. Initializing the level set surface by calculating the contrast gradient to estimate boundary partitions. 2. Iteratively optimizing segmentation curves until convergence of the energy function yields optimal results. 3. Utilizing segmentation outputs to calculate tumor morphology parameters (width, depth, and volume).

Postprocessing

The Regionprops function in Python was used to analyze the segmentation results, calculating tumor width (Maximum Feret Diameter) and depth (minor axis length), and vectorizing 2D slices for 3D reconstruction.

Statistical Analysis

Pearson’s correlation coefficient was used to assess the association between MSOT-derived and histopathology measurements. Margin of Error (MOE), defined as the difference between the two measurements, provided insights into algorithm robustness and consistency, with mean and standard deviation calculated to evaluate performance.

Results

1. Comparison of Tumor Width and Depth Measurements:

   • The correlation coefficients between MSOT and histopathology measurements for width and depth were 0.84 and 0.81 (p<0.0001), respectively, indicating strong correlation.
   • MOE analysis showed that 96% of depth data and 51% of width data deviated less than 1.5 mm from the gold standard.

2. Case Study and 3D Reconstruction:

   • In Case Study 36, the segmented 3D tumor reconstruction significantly reduced noise and calculated an overall tumor volume of 12.14 mm³.
   • Texture metrics derived from the Gray Level Co-occurrence Matrix (GLCM)—contrast (28.46), dissimilarity (1.99), homogeneity (0.64), etc.—provide potential insights for further research.

3. Time Efficiency:

   • The entire pipeline from imaging to data processing takes just 20 minutes, significantly improving efficiency compared to traditional biopsy and histopathological analyses that take several days.

Discussion and Significance

1. This study’s novel MSOT and segmentation approach overcomes traditional limitations in depth penetration and real-time monitoring capacity.
2. The non-invasive, real-time, 3D tumor evaluation facilitates precise preoperative mapping and clinical planning.
3. By capturing multidimensional tumor metrics (width, depth, and volume), this study enhances the comprehensiveness of pathology research.
4. Combining texture analysis with photoacoustic spectral data shows potential for exploring tumor microenvironments (e.g., angiogenesis, inflammation).

Prospects and Conclusion

This innovative study demonstrates the effectiveness of integrating MSOT with a robust level set segmentation algorithm, providing a user-friendly and precise method of tumor boundary measurement, while opening new perspectives in skin cancer imaging diagnostics. Future work will aim to combine machine learning models to optimize algorithm performance and expand application scenarios to non-pigmented tumors and complex clinical cases.

This breakthrough highlights the potential of photoacoustic imaging in dynamic tumor monitoring and establishes a solid foundation for personalized healthcare and rapid diagnostics. It marks a milestone in skin cancer detection technology, offering valuable directions for the future of precision medicine.