Classification and detection of basal cell carcinoma of the eyelid using convolutional neural networks and object recognition
Author: Christina Miller
Base Hospital / Institution: University hospital Ulm
Rapid fire oral presentation
Abstract ID: 24-473
Purpose
Basal cell carcinoms (BCCs) are the most common malignant periocular tumours. Early detection is crucial as these tumors spread locally by infiltration. Early forms of BCC may mimic other lesions and are often overlooked during a standardized slit lap ophthalmic examination. The aim of this study was therefore to establish an oculoplastic specialists trained neural network for differentiating BCC from other eyelid tumours based on clinical photographs; based upon the labeled model, a specialized object recognition model may be trained that enables detection and digital marking of BCCs on image frames allowing a video based alert of suspicious lesions during standardized ocular examination.
Methods
This is a retrospective study approved by local ethics committee. A database comprising 1210 photographs of 1237 excised and histologically confirmed tumours from 806 patients was created. Histopathological results were used to annotate images by experienced oculoplastic experts. The dataset was divided into training, validation and test dataset to train different neural networks. In a second step, the tumours were marked on images directly by computer vision using the YOLO algorithm, with a darknet-53 backbone serving as an object recognition model.
Results
The accuracy of detection of BCCs and differentiation from other tumour entities was 80% with a sensitivity of 95% and a specificity of 37%. The object recognition model was also able to identify tumours directly on images, with the majority of tumours correctly marked.
Conclusion
Our study shows that the detection of BCC is feasible using a neural network and that the differentiation from other eyelid tumours can be correctly performed based on photographs in four out of five tumours at first presentation. To avoid missing periocular lesions, the object detection model could be integrated into the video signal of a slit lamp, providing warning and digitally marking the tumour visibly.
Additional Authors
| First name | Last name | Base Hospital / Institution |
|---|---|---|
| Marie-Luise | Stock | University hospital Ulm |
| Armin | Wolf | University hospital Ulm |
| Christian | Wertheimer | University hospital Ulm |
| Lennart | Hartmann | University hospital Ulm |
