Eye Cancer Detection and Identification
Eye cancer is a rare malady, often challenging to identify at an earlier stage, and thus becomes immedicable if left undiagnosed. Various testing methods have yet been proposed in this domain; however, each method has its strengths and limitations. For example, detection using a gonioscopy lens can
2025-06-28 16:32:29 - Adil Khan
Eye Cancer Detection and Identification
Project Area of Specialization Artificial IntelligenceProject SummaryEye cancer is a rare malady, often challenging to identify at an earlier stage, and thus becomes immedicable if left undiagnosed. Various testing methods have yet been proposed in this domain; however, each method has its strengths and limitations. For example, detection using a gonioscopy lens can only be done if the tumor of considerable size. Indirect ophthalmoscope-based techniques are not effective as there exist human errors. High-energy sound waves-based ultrasound biomicroscopy (UBM) requires water-bath immersion with direct contact to the eye and has a longer image acquisition time. Biopsy, screening by removing a small part of cancerous tissues, is rather a painful procedure for eye melanoma analysis.
Artificial intelligence technologies have been proved to be a reliable tool in the field of medical engineering. Therefore, reflecting the significance of impeccable identification, our aim is to develop a medical-imaging based intelligent diagnosis system for eye cancer detection and identification. Our research work revolves around the efficient detection and classification of ocular cancers. Image processing and neural network-based classifiers will be implemented on the fundus images to automate the cancer evaluation and diagnosis process. Fundus images will be preprocessed and passed into deep convolutional neural network (CNN) architecture, where features will be extracted to classify eye malignancies. The proposed system’s results will then be validated using several reliable evaluation metrics, such as F-1 score, classification accuracy, confusion matrix, and logarithmic loss.
No such system is in use in Pakistan, and mostly the medical practitioners rely on conventional techniques i.e., the patients have to undergo a series of tests that have their shortcomings. Our project has the potential to extend public access to clinical diagnosis with reduced healthcare costs. The proposed system can be installed at infirmaries with optimized computer-assisted screening for cancer identification.
Project ObjectivesThe main aim of our project is to detect and classify the type of eye cancer based on machine learning techniques. This can be done by training a CNN model on the fundus images dataset. To test the proposed system in real-time, retinal images would be fed into the simulation software via a high-resolution lens interfaced to a camera.
The objectives of our project are:
- To evaluate different deep learning algorithms for better cancer detection.
- To develop a cost-effective and optimized eye cancer detection system.
- To make our ocular oncology units capable of utilizing artificial intelligence (AI) based diagnosis systemsÂ
Our project is based on three execution phases. First, the implementation of software, in which open access fundus datasets would be used to train the classification model. Second, the design and development of a camera for fundus photography. The third phase is the combination of the first two phases, i.e., the verification and testing of the trained model in real-time.
It would go through the following steps to reach the end goals of detection and identification.
- Several open access fundus datasets would be merged into a single large dataset. It would have 26 classes of eye disorders for identification.
- The dataset would undergo various preprocessing techniques to make it ready for the next phase.
- The processed data would be split into a training set and test set.
- Different classifiers would be developed for tumor identification using machine learning algorithms.
- A training set would be used to train the classifier. The training stage involves the k-fold cross-validation technique.
- A test set would be used to make predictions.
- Evaluation metrics would be used to compare the performance of each classifier.
- The model with top-of-the-range performance would be selected.
- An ocular imaging adapter would be used to capture fundus images in real-time.
- The images would be preprocessed by various image processing techniques.
- The processed images would be supplied as input to the pre-trained model.
- The model would detect and classify the type of eye cancer.
Medical diagnosis using artificial intelligence is an interesting field and is gaining popularity very rapidly. Our project aims to serve society clinically as well as economically. Conventional fundus screening tests are not only costly but take time and require skilled medics for observing the minor changes in surrounding nerves. On the other hand, deep learning-based intelligent systems can efficiently detect and identify the class of disorder. But obviously, this ability of automated systems could never take place of ophthalmologists, rather it would give them a streamlined diagnostic tool for cancer detection and identification.
The system would revolutionize the current practices if implemented successfully. Generally, when a fundus photograph is taken, an ophthalmologist analyzes it to identify the type of cancer. This task of inferring the type of malignancy, by inspecting the retinal nerves, requires high proficiency and observation skills. An artificial intelligence-based system would empower ophthalmologists by providing them a smart diagnosis tool. However, the integrity and reliability of our AI-based framework would be validated before substituting current practices, in order to ensure that the services are risk-free.
The proposed detection and identification system, inspired by the human multilayered neuronal system, is aimed to be capable of minimizing workload by functioning as a primary filter before the patient is examined by a doctor. In this domain, our proposed model would automatically pinpoint, analyze, and identify the pathological data in a comprehensive, rapid and non-invasive manner. It tends to transform individual healthcare in near future.
Technical Details of Final DeliverableFinal deliverable includes the following modules,
- Deep-learning based systemized intelligent program with optimal results
- A 3D printed ocular imaging adapter
- Real-time implementation and validation results
| Item Name | Type | No. of Units | Per Unit Cost (in Rs) | Total (in Rs) |
|---|---|---|---|---|
| Total in (Rs) | 65740 | |||
| 20-Diopter condensing lens (50 mm) | Equipment | 1 | 56000 | 56000 |
| 3D printed accessories | Equipment | 8 | 1200 | 9600 |
| M3 Bolts and nuts (15 mm) | Miscellaneous | 8 | 15 | 120 |
| M8 Bolt and nuts (25 mm) | Miscellaneous | 1 | 20 | 20 |