Science and technology

In a groundbreaking doctoral study, artificial intelligence (AI) has proven its ability to accurately distinguish malignant skin lesions from benign ones using hyperspectral imaging (HSI). Prior to AI analysis, the skin lesions were imaged using hyperspectral cameras.

Skin cancers are among the most common cancers, and their incidence is steadily increasing in Finland and worldwide.

Several non-invasive imaging methods have been developed for identifying skin cancers, but they often require specialized training for image analysis. Medical licentiate Vivian Lindholm conducted research to explore whether hyperspectral imaging (HSI) combined with AI analysis could determine which skin lesions are malignant and which are benign.

Hyperspectral imaging holds promise as a tool for diagnosing skin cancer. In practice, a hyperspectral camera captures a broader range of information about different light wavelengths compared to a regular camera. In Lindholm's research, the camera provided unique spectral data, akin to a "fingerprint," for each skin lesion. This spectral data was combined with a 3D model depicting the surface structure of the skin lesion. AI analysis, utilizing neural networks, processed the data and provided a diagnosis for each skin lesion.

Lindholm assessed the accuracy of this method in cases where the goal was to differentiate melanoma from benign pigmented nevi, keratinocyte skin cancer from benign raised moles, and precancerous skin changes (Bowen's disease) from benign seborrheic keratoses. The study included 172 skin lesions.

The research findings indicate that AI analysis following hyperspectral imaging effectively and accurately distinguishes malignant skin lesions from benign ones. It achieved a 95% accuracy and sensitivity for melanoma, 85-100% sensitivity, and 92-100% accuracy for keratinocyte skin cancer, and 84% sensitivity and 94% accuracy for Bowen's disease.

Additionally, this analysis provides a mapped image of the skin lesions, allowing for more precise delineation of lesion borders, particularly useful before surgical treatment.

"This method appears promising as an automated means of identifying skin cancers, which could assist doctors in pinpointing skin lesions that require treatment more accurately, reducing the need for unnecessary biopsies and enabling the early detection of skin cancers. However, further research is needed, including large datasets encompassing all types of skin lesions, to validate and refine the method," says Vivian Lindholm, a doctoral researcher at the University of Helsinki and specialist physician at Helsinki University Hospital's Inflammation Center.

Research Also Explores Treatment for Precancerous Skin Changes

Preventing squamous cell carcinoma with photodynamic therapy (PDT) for precancerous skin changes, known as actinic keratoses, is effective but often painful. Daylight PDT, which utilizes natural sunlight, is nearly painless but less effective for thick lesions. Recently, indoor simulated daylight PDT has been introduced as an alternative.

Lindholm evaluated the effectiveness of two new laser-assisted PDT treatments for actinic keratoses on the head. The study included 115 patients.

The most significant benefit was observed when using fractional laser before simulated or natural daylight PDT. Fractional laser creates tiny holes in the skin's surface, significantly enhancing the efficacy of daylight PDT: 86% of the lesions improved with this combined treatment, while using daylight PDT alone resulted in a 70% improvement.

"Furthermore, this combination performed excellently for thick lesions, surpassing the effectiveness of standard PDT and being well-tolerated. It could potentially become the primary choice for patients with extensive or thick actinic keratoses. However, additional research is needed to optimize laser settings, enhancing both treatment efficacy and tolerance," Lindholm added.

Fact: Hyperspectral Imaging Captures Extensive Skin Data

  • A conventional photograph pixel contains information on the intensity of three different wavelengths, whereas hyperspectral cameras record data on dozens of different wavelengths per pixel.
  • Information on light intensity at various wavelengths provides graphic spectral data, offering more detailed insights into the composition of the imaged skin area.
  • In skin imaging, the spectrum primarily depends on chromophores, such as skin melanin, hemoglobin, bilirubin, water, and fat content. Thus, the spectrum is unique for each skin lesion.
  • Due to the vast amount of data in hyperspectral imaging, recent studies have leveraged AI methods for analyzing hyperspectral camera results.