1. Use

• professional medical imaging and analysis test non-invasive diagnostic aid

• indicates probability of melanoma in pigmented skin lesions (moles)

• intended for use by authorised (trained) operators (medical professional) as adjunct screening technology in clinical pathway of the management of suspect lesions
  

2. On-device AI

• supplied as proprietary software on a single-application iOS device

• secure image capture with all other functions disabled

• nomela® test and analysis by machine-learning AI made directly on the device

• no requirement to internet connectivity for test and analysis completion

• not an on-line web-based application
  

3. User management, Quality and IT security

• authorised user management; site and date/time information, capture of minimal patient identification (NHS/CHI/hospital number acquired by barcode recognition in device), region & close-up images (for unequivocal identification of lesions), multiple lesion test capacity; e-consent included if required

• no requirement for dermatoscopic images

• QMS ISO13485 certified with regular audit by Notified Body (BSi); IT security certified by Cyber Security Plus and NHS DSP Toolkit; MDR Class IIa in process
  

4. Analysis technology

• nomela® analysis subject to:
  i) no contra-indications (operator confirms)
  ii) distinct edge obtained (automated)
  iii) size of lesion (not less than 5mm diameter including calculation by the test)

• analysis technology employs proprietary automated lesion edge detection and on-device machine-learning AI
  

5. Result and Reporting

• only a few minutes required to complete the image capture, analysis and report

• Result shown on the device screen as Low or High Chance of melanoma

• Report shows user information, NHS/CHI/hospital number, high-quality images and nomela® analysis result(s)

• available for local printing or upload by cellular (4G/5G) connection via AWS serverless server to secure (nhs.net) email or to allow access by authorised ePR

The nomela® AI architecture

The nomela® AI architecture (nomela® v6) is composed of two key components:
(i) a re-trained convolutional body model, adept at transforming input model images into a sequence of distinctive “features”, and
(ii) a fully connected head model responsible for making the ultimate classification decision. The training process exclusively focuses on fine-tuning the head model. More than 40 candidate body models were evaluated and compared to optimise performance.

Data description

The nomela® v6 AI system has been trained and validated rigorously on a data subset comprising 679 lesions (107 melanoma, 572 not-melanoma). The system has been subsequently tested on an independent subset of 450 lesions (70 melanoma, 380 not-melanoma). All lesion diagnoses had been confirmed by histopathology. For pre-processing, an automated technique has been, and is, used to delineate the lesions in a deterministic manner.

Results

The performance of the nomela® v6 AI system at 90% sensitivity is a specificity of 66% with a ROC AUC of 0.85. This is non-inferior to 4 human (dermatologist) experts assessing the same testing subset of images.