Smartphone-Powered Deep Learning In Action: Early Detection Of Malignant Transformation In Conjunctival Nevi

To evaluate a novel deep learning model for noninvasive diagnosis purposes in the detection of size and borders of conjunctival pigmented nevi, aiming to prevent malignant transformation.

The patient group consisted of 508 nevi and 538 normal eye images captured with an A71 Samsung mobile phone, randomized by sex; however, the age range was between 20 and 70 years old. 

The patient group consisted of 508 nevi and 538 normal eye images captured with an A71 Samsung mobile phone, randomized by sex; however, the age range was between 20 and 70 years old. The UNet model was developed with a customized encoder and decoder, utilizing 80% of the dataset for training with nevi exact border annotation labeling, while the remaining 20% was reserved for testing.

The average Weighted Dice coefficient score for masks generated by the UNet model was 98.48% on the test dataset for recognizing nevi regions, detecting exact borders, and estimating nevi size. The model demonstrated remarkable performance in segmenting nevi regions, highlighting its ability, performance, and robustness for our segmentation task. The close alignment between human labeling and the masked border of nevi in this model indicates its capability to generate accurate masks that capture the shape of nevi images.

The model is capable of accurately detecting the size and border of conjunctival nevi using images captured through a mobile phone camera. This capability can facilitate alongside the universal movement in AI-based diagnostic and monitoring healthcare mobile applications, potentially serving as a lifesaver in the malignant transformation during its early stages with the utilization of this novel deep learning model.