1243P - Ocular surface squamous neoplasia early diagnosis using an AI-empowered autofluorescence multispectral imaging technique

Background

Ocular surface squamous neoplasia (OSSN) is a common eye surface neoplastic disease ranging from non-invasive conjunctival intraepithelial neoplasia (CIN) to invasive forms of squamous cell carcinoma (SCC) that accounts for approximately 10% of ocular surface lesions. Early detection of OSSN is crucial for effective treatment. However, clinical symptoms of OSSN vary, which can result in delayed or inappropriate diagnosis without appropriate diagnostic methods. Currently, the diagnosis of OSSN relies on clinical suspicion confirmed by impression cytology or biopsy, which are invasive and may not detect small lesions. This study aims to introduce an artificial intelligence data analysis framework to empower a newly designed autofluorescence hyperspectral imaging technology developed in our group for OSSN detection.

Methods

We recruited 20 patients who had previously been diagnosed with OSSN. Biopsy specimens were collected from the patients and reprocessed without staining to obtain autofluorescence microscopy images using 25 spectral channels, each with a distinctive excitation-emission wavelength. The spectral images were pre-processed to remove unwanted random noise and then spectral features were extracted from the epithelial section of tissue images. Further, the spectral features were used to form an artificial feature image patch to train a convolutional neural network for image classification.

Results

The technique was evaluated against pathological assessment, and results showed the multispectral analysis was able to detect OSSN in all tested patients (P-value < 0.05). The trained classifier on OSSN and normal sections was found to have great performance (AUC > 83%). The maps produced by the analysis were in close agreement with margins observed in H&E sections. The fully automated diagnostic method based on AI methodology produced maps of the neoplastic-non neoplastic interface that could be generated in quasi-real time and used for intraoperative assessment.

Conclusions

The study demonstrated the feasibility of using AI empowered hyperspectral autofluorescence imaging to detect OSSN and the potential for the technique to be used as a diagnostic tool in future clinical applications.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Cancer Institute NSW.

Disclosure

All authors have declared no conflicts of interest.