Classification system for skin lesions, the more detailed the better

Skin cancer is one of the most common cancers in the world. Late-stage skin cancers spreads to internal organs and become fatal. Early-stage skin cancer can be cured with a high survival rate, while the 5-year survival rate for skin cancer is extremely low. Therefore, early detection is the key in fighting skin cancer.

The use of artificial intelligence (AI) in medical diagnosis has become very popular for many decades. Thousands of scientific papers have already shown that AI can support dermatologists in detecting malignant skin lesions. Many AI solutions have been developed for skin cancer detection. The famous international ISIC challenges provides tens of thousands dermoscopic images of skin lesions to competitors to build AI algorithms that can classify skin lesions into several classes, including malignant melanoma, the deadliest skin cancer.

Supported by iToBoS, one of the biggest European projects for fighting skin cancer, Torus Actions has developed a novel AI solution for skin cancer detection called Skin Cancer AI. This Skin Cancer AI allows to classify clinical and dermoscopic images of skin lesions into more than 100 categories and estimate the cancer risk level.

Back to the early development of Skin Cancer AI, the starting point to build the classification system was the one of ISIC challenge 2019 which contains 9 classes: (MEL) Melanoma, (NV) Melanocytic nevus, (BCC) Basal cell carcinoma, (AK) Actinic keratosis, (BKL) Benign keratosis (solar lentigo / seborrheic keratosis / lichen planus-like keratosis), (DF) Dermatofibroma, (VASC) Vascular lesion, (SCC) Squamous cell carcinoma, and (UNK) None of the above.

The ISIC 2019 classification system covers the most common malignant (MEL, BCC, AK, SCC) and benign (NV, BKL, VASC, DF) classes of skin lesions. It is easy to annotate for the doctors and more informative than the binary “malignant vs benign” classification of ISIC challenge 2020.

However, this classification system still has some weaknesses such as:

  • Ambiguity in the classification. For example, clear cell acanthoma (CCA) is a benign keratinocytic lesion but is not in the list of subclasses of BKL. Should it belong to BKL or UNK? Many rare-but-not-too-rare types of skin lesions, both benign and malignant, are missing in the classification system, hence they all become UNK. Examples include Merkel cell carcinoma (MCC), Kaposi sarcoma (KS), dermatofibrosarcoma protuberans (DFSP), sebaceous gland hyperplasia (SGH), etc.
  • Indeed, Bowen disease (BW) lesions were included in AKIEC (assimilated to the same class as AK) in the ISIC 2018 classification system. However, in ISIC 2019, it is considered as a part of SCC. Therefore, the ISIC classification system is not backward compatible.
  • Not enough information for determining the cancer risk level. Indeed, lesions in the same class in ISIC classification system may have very different cancer risk levels. For example, pyogenic granuloma (PG) and cherry angioma are both in VASC. However, a PG may have a much higher cancer risk since it may mimic malignant melanoma or other malignant vascular lesions. Similarly, in NV class, a spitzoid nevus or a dysplastic nevus will also carry a much higher risk than a common congenital nevus.

By making a throughout research about skin lesion classification and with the help of our advisor dermatologists, we have built a new classification system with 10 classes and more than 100 subclasses. The ten classes are: (MEL) Malignant melanoma, (BCC) Basal cell carcinoma, (EPI) Epidermal pre-malignant and malignant tumours, (MALO) Other cutaneous malignant tumours, (NV) Melanocytic nevus and melanosis, (DF) Dermatofibroma, (BAL) Benign adnexal lesions, (BKL) Benign keratinocytic legions and lentigines, (VASC) Benign vascular malformations and haemorrhages, and (BENO) Other benign lesions and non-tumours.

Our classification system addresses the weaknesses of the ISIC 2019 classification system and is much more detailed and informative. In particular, the “UNK” class of ISIC 2019 is divided into three separate classes, with useful information:

  • MALO: malignant lesions other than MEL, BCC, AK, BW, SCC, and other-epidermal malignant lesions. This class includes important subclasses missed by the ISIC 2019 classification system such as KS, MCC and DFPS.
  • BAL: benign adnexal lesions such as SGH and poromas.
  • BENO: other benign lesions and non-tumors such as lichen planus and lupus erythematosus.

Besides, the two classes AK and SCC of ISIC 2019 classification system become subclasses of EPI which also contains BW and other epidermal pre-malignant tumours.

The detailed information contained in Torus classification system will make a more comprehensible diagnosis of skin lesions. Moreover, it allows for a much more precise estimation of the cancer risk level for skin lesions.