top of page
  • Aurel Iuga

Artificial‌ ‌intelligence-enabled‌ ‌diagnostic‌ ‌support‌ ‌for‌ ‌cutaneous‌ ‌melanoma

In Romania, every year, more than 400 people die prematurely due to skin melanoma.[1] If detected early enough, skin melanoma can be cured by a simple surgical excision of the skin. In addition, the 5-year survival rate for cutaneous melanoma ranges from 99% for localized diseases (stages 0, I and II) to 65% for regional disease (stage III), to only 25% for distant disease (stage IV) (Figure 1).[2]

Fig 1. 5-year survival rate for cutaneous malignant melanoma

The diagnosis of cutaneous melanoma is, unfortunately, made at a much more advanced stage of the disease in Romania than in other EU countries. For example, in Sweden, more than 92% of diagnoses are made when the disease is localized [3], while in Romania only 34% of cases are diagnosed at this stage, the rest of the cases being composed of patients diagnosed with regional or distant disease (Figure 2).[4] Also, while Western European countries have a 5-year survival rate of about 90%, Romania unfortunately has a 5-year survival rate of only 50-60% for patients with cutaneous melanoma.[4,5]

Fig 2. Cutaneous melanoma, the stage of the disease at diagnosis in Sweden vs. Romania

Globally, screening programs have been successfully implemented for at least several types of cancer.[6,7] In these situations, there are diagnostic technologies and treatment options that make the benefits of these screening programs, at the population level, outweigh their disadvantages. Examples of cancers in which screening programs have been shown to be effective at the population level are lung, breast, cervical, colon or skin cancers. For cutaneous melanoma, early detection makes a clear difference between life and death.

Advances in computing infrastructure, the development of artificial intelligence (AI) algorithms, and the accumulation of large data sets have in recent years propelled image processing and analysis using AI.[8] In medicine, such algorithms are now able to classify medical images and provide diagnostic support in a multitude of clinical scenarios such as diagnostic radiology, retinal imaging, histopathology, etc.[9] Explaining further, Esteva et al. (2021) prepared a detailed report of the computerized analysis of imaging data in medicine, including its numerous applications to date.[8]

The field of dermatology has also made progress in the classification of skin photographs, as well as dermatoscopic imaging, with a performance of AI models equal to or sometimes surpassing human specialists.[10, 11, 12] Computer-assisted interpretation of skin images has the opportunity to improve upon at least some of the current problems in clinical practice by improving test performance (accuracy / sensitivity, recall / positive predictive value). Diagnosis assisted by artificial intelligence has other advantages compared to the traditional approach. It provides almost real-time preliminary test results, which can help improve patient adherence to follow-up recommendations. This is important intuitively, especially for vulnerable populations.

Artificial intelligence also has the potential to reduce the total costs of the screening program by facilitating the triage of cases with a low probability of malignancy (probability generated by the artificial intelligence algorithm) and/ or by eliminating the need for additional confirmation by a dermatologist. By lowering costs and improving patients' access to immediate test results, artificial intelligence enabled screening programs for skin melanoma have the potential to reduce some disparities in healthcare. More effective implementation of screening programs, with wider patient access, will have a positive impact on health outcomes. However, additional studies are needed in Romania to determine the overall impact of skin cancer screening programs on overall medical costs.

With a relatively small investment, taking into account the benefits of such a program, skin melanoma screening programs can be developed to save the lives of many of these patients in Romania. Therefore, we propose the development of a software tool that will facilitate the screening of skin melanoma and provide diagnostic assistance. This tool will capitalize on the analysis of high-resolution photographs of skin lesions, as well as dermatoscopic imaging. To this end, we report here the preliminary development of diagnostic automation algorithms for this purpose. The financing of this project will allow us to complete these algorithms necessary to implement a screening program.

As part of this proposal, we also aim to implement a pilot clinical program in family medicine practices, as well as dermatology practices. This pilot program will assess the operational feasibility as well as the impact on the clinical outcome of such a screening program. All these technical resources and the results of the pilot project can be used later to develop a clinical screening program at national level in Romania.


1. Cancer today.

2. Survival Rates for Melanoma Skin Cancer.

3. Rockberg, J. et al. Epidemiology of cutaneous melanoma in Sweden-Stage-specific survival and recurrence rate. International journal of cancer 139, (2016).

4. Rotaru, M., Jitian, CR & Iancu, GM A 10-year retrospective study of melanoma stage at diagnosis in the academic emergency hospital of Sibiu county. Oncol. Lett. 17, 4145–4148 (2019).

5. Forsea, A.-M. Melanoma Epidemiology and Early Detection in Europe: Diversity and Disparities. Dermatol Pract Concept 10, e2020033 (2020).

6. Screening Guidelines and Other Resources. (2021).

7. Vrdoljak, E. et al. Cancer Control in Central and Eastern Europe: Current Situation and Recommendations for Improvement. The Oncologist vol 21 1183–1190 (2016).

8. Esteva, A. et al. Deep learning-enabled medical computer vision. NPJ Digit Med 4, 5 (2021).

9. Ting, DSW et al. AI for medical imaging goes deep. Nat. Med. 24, 539–540 (2018).

10. Brinker, TJ et al. Deep learning outperformed 136 of 157 dermatologists in a head-to-head dermoscopic melanoma image classification task. Eur. J. Cancer 113, 47–54 (2019).

11. Valle, E. et al. Data, depth, and design: Learning reliable models for skin lesion analysis. Neurocomputing Vol 383 303–313 (2020).

12. Esteva, A. et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature vol. 542 115–118 (2017).

bottom of page