View Booklet Team Phoenix Project report flipbook.
P R E P A R E D B Y CoE – AI, RMKEC A C A D E M I C Y E A R 2023 – 2024 P R O G R A M M E Hack to Hire 2024 C E N T R E O F E X C E L L E N C E Artificial Intelligence Industry Problem Statements & Use Cases Knowledge Partner – Cognizant Real-world industry challenges, framed for AI-driven solutions by students and faculty. PROJECT REPORT DETECTING TUMOURS FROM IMAGING Academic Year 2023-2024 R.M.K. ENGINEERING COLLEGE (An Autonomous Institution) RSM NAGAR, KAVARAIPETTAI – 601 206 Affiliated to Anna University, Chennai / Approved by AICTE, New Delhi / Accredited by NAAC with A+ Grade / ISO 21001:2018 Certified Institution / All the eligible UG Programs are Accredited by NBA, New Delhi
R.M.K. ENGINEERING COLLEGE (An Autonomous Institution) R.S.M NAGAR, KAVARAIPETTAI – 601 206. Gummidipoondi Taluk, Thiruvallur District, Tamil Nadu. Department of Artificial Intelligence and Data Science PROJECT REPORT DETECTING TUMOURS FROM IMAGING Cognizant Hack to Hire 2024 · Academic Year 2023-2024 Team Team 7 – Team Phoenix Faculty mentors Dr. Sandra Johnson Students Shaik Eshrath (TL) – 111720201042, Modepalli Venkata Sai Chandana – 111720201024 Cognizant mentors Mr. Brijmohan Jakhmola, Ms. Payal Mangal Abstract Interpreting MRI scans to detect brain tumours needs expertise, takes time and is prone to human error. This project detects, segments and classifies brain tumours from MRI scans and generates a downloadable scan report. A 3D U-Net segments the tumour region at pixel level (96.2% accuracy, mean IoU 0.585), and a CNN classifies the tumour as glioma, meningioma, pituitary or no tumour (88% accuracy). The tumour area and volume are used to estimate its stage. The models are deployed as Amazon SageMaker endpoints, with uploaded scans stored in S3, behind a Django web application. Problem Statement Improve tumour detection in medical imaging using imaging technology, machine learning and clinical expertise: a user-friendly interface to upload scans, pre-processing, models for tumour segmentation and classification, analysis of tumour characteristics for treatment planning, and continuous improvement from feedback. Objectives • Segment the exact tumour region from brain MRI scans. • Classify the tumour type: glioma, meningioma, pituitary or no tumour. • Estimate tumour area, volume and stage. • Generate a downloadable scan report through a web interface.
System Architecture Fig 1. System architecture (team’s final presentation) Methodology • Datasets: Kaggle brain MRI datasets, with 2,870 classification images (glioma 886, meningioma 823, pituitary 834, no tumour 327), augmented to 4,780 images at 224 × 224; BraTS-style multi- modal MRI (FLAIR, T1, T1ce, T2) for segmentation. • Augmentation: resizing, flipping, shearing, rotation up to 10°, 5% noise and blur up to 1.5 px. • Segmentation: 3D U-Net (encoder–decoder) compared with 2D U-Net and YOLOv8; 3D U-Net chosen because it preserves spatial information in all three dimensions. • Classification: CNN compared with a Vision Transformer; the CNN was chosen for its higher accuracy. • Staging: tumour area mapped to stage 1 (0–7 mm²), 2 (7–14), 3 (14–21) or 4 (above 21). • Deployment: models deployed as Amazon SageMaker endpoints; S3 stores uploaded scans; Django front end with report download. Technology Stack Layer Tools Deep learning Python, TensorFlow / Keras, PyTorch, 3D and 2D U-Net, YOLOv8, CNN, Vision Transformer Cloud Amazon SageMaker endpoints, Amazon S3; Google Colab GPU for training Application Django, HTML, CSS, JavaScript; downloadable scan report
Results Segmentation and classification models compared (team’s evaluation-metrics slide). Task Model Accuracy Other metric Segmentation U-Net 3D (selected) 96.2% Mean IoU 0.585 Segmentation U-Net 2D 97.0% Mean IoU 0.570 Segmentation YOLOv8 89.9% – Classification CNN (selected) 88% Weighted F1 0.87 Classification Vision Transformer 80% – CNN per-class F1: glioma 0.87, meningioma 0.74, no tumour 0.93, pituitary 0.94 (1,311 test images). Fig 2. 3D U-Net results – test image, label, prediction and training curves (team’s presentation) Outcome and SDG Mapping The system was reviewed by Cognizant and faculty mentors on 12 July and 25 July 2023 (demo), and the models were deployed on Amazon SageMaker by 31 July 2023. The project maps to SDG 3 – Good Health and Well-being. Conclusion The system turns an MRI scan into a segmented tumour region, tumour type, size and stage within minutes, compared with about a week for a traditional radiology report. Future work includes liver, breast and skin tumours, cloud hosting of the full application, and a digital twin of the brain for treatment planning.
Gartner Hype Cycle 2023: Technology Position The Detecting Tumours from Imaging project was done in July 2023 under Cognizant Hack to Hire 2024. Its main technology has been compared with the Gartner Hype Cycle for Artificial Intelligence, 2023, released in August 2023, which was the edition in use at that time. On that Hype Cycle the project comes closest to Computer vision. Gartner's benefit rating is left out, since the use case came from Cognizant and the business value of the solution was not assessed. Hype Cycle attribute (2023) Computer vision Hype Cycle phase Slope of Enlightenment, moving towards the Plateau of Productivity Maturity Early mainstream* Market penetration 20–50% of target audience* Time to mainstream adoption Less than 2 years* Related 2023 profiles Cloud AI services (Early mainstream, 5–20%, moving towards the Plateau), since the models were deployed on Amazon SageMaker Meaning of the Attributes Attribute What it describes Gartner scale Computer vision in 2023 Maturity How far the technology itself has developed Embryonic, Emerging, Adolescent, Early mainstream, Mature mainstream, Legacy Early mainstream. The technology is proven, products are widely available and adoption is growing quickly. Market penetration The share of the intended users who already use it <1%, 1–5%, 5–20%, 20– 50%, >50% of the target audience 20–50%. Between a fifth and a half of target organisations already use it. Time to mainstream adoption The number of years before it reaches the Plateau of Productivity, where it is widely used and gives dependable results <2, 2–5, 5–10, >10 years, or obsolete before reaching the plateau Less than 2 years. Routine use was expected by about 2025.
* Gartner's 2023 report places this technology on the Slope of Enlightenment, moving towards the Plateau of Productivity, and describes it as commoditised. The maturity, penetration and time-to-mainstream values marked * have been assigned on Gartner's scale from that 2023 position. Computer vision was already a proven technology when Team Phoenix took up this use case in July 2023, and Gartner placed it close to mainstream use. The team applied it to detecting tumours from imaging. Gartner Hype Cycle Graph Figure: Position of Team Phoenix on the Gartner Hype Cycle for AI, 2023 Reading the Graph The red dot marks Team Phoenix. It sits on the Slope of Enlightenment because computer vision was at that stage in 2023. On the Slope of Enlightenment the benefits are well understood, improved products are available and the technology is moving into routine use. The grey dots are other AI technologies from the same Hype Cycle, shown for comparison. How the Project Relates to the Profile • The team segments brain tumours from MRI scans with a 3D U-Net (96.2% accuracy, mean IoU 0.585) and also trained CNN and Vision Transformer classifiers. All of these fall under computer vision. • The models were deployed on Amazon SageMaker, which relates to Gartner's cloud AI services profile. • Both computer vision and cloud AI services were moving towards the Plateau in 2023, so the team could concentrate on accuracy for the clinical task.
Annexure: Final Project Presentation Slide 1 Slide 2 Slide 3 Slide 4 Slide 5 Slide 6
Slide 7 Slide 8 Slide 9 Slide 10 Slide 11 Slide 12 Slide 13 Slide 14
Slide 15 Slide 16 Slide 17 Slide 18 Slide 19 Slide 20 Slide 21 Slide 22
Slide 23 Slide 24 Slide 25 Slide 26 Slide 27 Slide 28 Slide 29 Slide 30
Slide 31 Slide 32 Slide 33 Slide 34 Slide 35 Slide 36 Slide 37 Slide 38
Slide 39 Slide 40 Slide 41 Slide 42 Slide 43 Evidence Source code https://github.com/Eshrath027/Detectiong-tumors-from-imaging Demo video https://drive.google.com/file/d/1J6NknLZ0l8l_waHCnBj_xijhyP-YxydG/view Final presentation https://drive.google.com/file/d/10bggPBTh-e9E0Mgj85zGEsz7jtGKd_vL/view