Ir para o conteúdo principal

Utility and Reliability of a Deep Learning Algorithm as a diagnosis support tool in head & neck non-melanoma skin malignancies

· 3 minutos de leitura
European Archives of Oto-Rhino-Laryngology
European Archives of Oto-Rhino-Laryngology
Peer-reviewed journal

Published in European Archives of Oto-Rhino-Laryngology, this prospective study evaluated Legit.Health as a diagnosis support tool in the head and neck surgery clinic of Donostia University Hospital, on the two most common skin cancers: basal cell carcinoma (BCC) and cutaneous squamous cell carcinoma (cSCC).

Medela A, Sabater A, Hernández Montilla I, Mac Carthy T, Aguilar A, Chiesa-Estomba CM. The utility and reliability of a deep learning algorithm as a diagnosis support tool in head & neck non-melanoma skin malignancies. Eur Arch Otorhinolaryngol. 2025;282(3):1585-1592. https://doi.org/10.1007/s00405-024-08951-z

Conclusions​

Legit.Health placed basal cell carcinoma and squamous cell carcinoma among its five most likely diagnoses in the great majority of cases, and told malignant from benign lesions apart with an excellent balance of sensitivity and specificity. It did so both on the lesions photographed in the clinic and on a larger, independent test set of the same conditions.

The authors conclude that Legit.Health has real value as an aid for the early detection and treatment of non-melanoma skin cancer, and see promise for its use in primary care, dermatology and head and neck surgery.

Summary​

  • Journal: European Archives of Oto-Rhino-Laryngology (peer-reviewed)
  • Design: Prospective, longitudinal, non-randomised
  • Setting: Head and neck surgery outpatient clinic, Donostia University Hospital (San Sebastián, Spain)
  • Study period: June to December 2021
  • Patients: 135 adults referred to the clinic with suspicious skin lesions
  • Lesions: 54 basal cell carcinomas, 54 squamous cell carcinomas and 27 benign lesions (seborrheic keratosis, actinic keratosis and actinic cheilitis)
  • Reference standard: Histopathology, after biopsy or surgical excision

Design​

Each lesion was photographed with a smartphone at 10 cm, under the clinic's ordinary lighting and without zoom or flash, and analysed by Legit.Health. The study assessed two outputs against histopathology: the ranked list of most likely diagnoses, and the malignancy estimation, evaluated with ROC analysis. The same analysis was repeated on 3,821 images of the same five conditions drawn from an independent test set, to check that the results held beyond a single clinic.

Results​

Both cancers were consistently ranked among the five most likely diagnoses, and basal cell carcinoma was most often the top-ranked one. Benign lesions were identified within the five most likely diagnoses in every case.

Benign and malignant lesions of the head and neck often look alike, seborrheic keratosis and squamous cell carcinoma in particular. Where that happened, Legit.Health kept both in its list rather than dropping the cancer, which favours sensitivity, the property that matters most when the question is whether a lesion could be malignant.

The full figures, per condition and per dataset, are in the paper above. As a next step, the authors propose evaluating Legit.Health with other cameras and in further settings, such as primary care.