Creating and sharing knowledge for telecommunications

IT Leiria innovating in cancer desease detection by using light field technology


by IT on 24-05-2023
Project Snapshot PlenoISLA project Processing, Analysis, and Coding of Audio and Visual Information SDG3 Good health and well-being
...

Catching up with...

 

Sérgio Faria

 

Currently, between 2 and 3 million non-melanoma skin tumors and around 132.000 melanoma skin cancers are detected per year worldwide. Their incidence has increased over the past decades, among them, melanoma is the most invasive skin cancer and with the highest risk of death. However, it is highly curable if detected in an early stage. Melanoma is a form of skin cancer that arises when pigment-producing cells (melanocytes) mutate and become cancerous. As it may be originated from another type of lesion, like nevus, it is quite important to identify these lesions as soon as possible.

Under the PlenoISLA project, the IT Leiria-branch research team investigated computational methods for skin lesion diagnosis to assist dermatologists and support expert systems. To this end, equipment was developed based on a camera with light field technology.  Using cameras with 42-megapixel sensors and a proper array of micro-lenses, this system allows the representation of skin lesions in 3D, facilitating the extraction of new features, beyond 2D.  The 3D information extracted from the light fields allowed to improve the performance of computer vision and machine learning algorithms in the evaluation of the cutaneous lesion, judging, not only if it is a benign or malignant cancer, but also the identification of other types of skin lesions.

By exploiting both 2D and 3D characteristics of the skin lesion surface, the proposed approach advances beyond commonly used color features of dermoscopic images. Two competing classification methods are exploited, namely Multiple Instance Learning (MIL) and Deep Learning (DL), which are combined using an uncertainty-aware decision function. The DL method performs classification resorting to RGB data, while MIL performs 3D feature extraction, selects the most significant set, and performs classification at two different learning instances. The novel aspects of this work include DL uncertainty evaluation mechanisms along with MIL to train a robust ensemble classifier, in addition to the use of dense light fields for skin lesion classification.

Despite the large class imbalance (often present in medical image datasets), the ensemble model achieves cross-validated melanoma classification accuracy of 84.00% when trained against nevus lesions, and 90.82% accuracy when discriminating against all present lesion types. The results show that, in the absence of discriminative 2D characteristics, the 3D surface provides redeeming results, demonstrating that existing methods can benefit from the proposed method by looking beyond 2D image characteristics.

This collaboration with the Department of Dermatology of Centro Hospitalar de Leiria allowed the creation of the first dataset of skin lesion light fields in the world (SKINL2), which is publicly available for the research community.

The outcomes of this research are expected to impact and benefit all the parties in the dermatological area, from researchers that will have a new type of data to develop new algorithms and tools, to users who will benefit from the technology, and practitioners that may increase the reliability of their diagnosis.

 

 

Listen to the podcast on our Youtube channel and Spotify!

 

Find out more on this project:


https://www.it.pt/Projects/Index/4583
SHARE: