Internship Report 6th Semester CSIR-CSIO.pptx

Medical Image Analysis
PPt for internship evaluation
Machine Learning
Internship semeter

1. ~ Research Internship Evaluation ~ CAD SYSTEM FOR BENIGN LIVER LESIONS USING GLCM DESCRIPTORS AND SVM CLASSIFIER Jyoti Dixit 21103013 Industry Mentor Dr. Jitendra Virmani College Mentor Dr. Padmavati

2. 2 RESEARCH ACCEPTED The current study titled as “CAD SYSTEM FOR BENIGN FOCAL LIVER LESIONS USING GLCM DESCRIPTORS AND SVM CLASSIFIER” has been accepted for inclusion in upcoming Book Series “AI and ML techniques in Image Processing and Object Detection”.  

3. Problem Statement To address the clinical challenge of accurately differentiating between FNH and HEM using B-mode ultrasound images, this study aims to develop and evaluate a machine learning-based CAD system. This study aims to address the clinical challenge of accurately differentiating between FNH and HEM ( using B-mode ultrasound images by developing and evaluating a machine learning-based computer-aided diagnostic (CAD) system. 3

4. PROJECT overvIEW STAGE 01 Understanding about the Basics of Digital Image Processing Understanding Fundamental working of MATLAB software. STAGE 02 Dataset understanding through research work. Surfed through Medical Image Databases STAGE 03 Dataset collection, Pre-processing & Literature Review of researches. STAGE 04 Feature Extraction Classifier Implementation ​ STAGE 05 Results analysis Conclusion of Research Acceptance of the CLASSIFICATION Disease classification and Analysis Feature extraction Texture Analysis Lesion Detection Detection

5. Presentation title 5 WHY CAD system ? Importance of early and accurate diagnosis of Benign Liver Disease Objective To prevent from further biopsies. Prevent last stage disease at the very early stage of Diagnosis. To classify benign focal liver lesions using Ultrasound Images and avoid unnecessary biopsies. Second opinion to Radiologists confirming their diagnosis.

6. Presentation title 6 Why liver ? M ost important tissue in human body 500 vital functions stores vitamins and carbohydrates, produces bile (a digestive fluid ), cleanses poisons from the blood. Remarkably, the liver also has the ability to repair itself . Why ULTRASOUND ? First examination of characterization of FLLs. Real time imaging capabilities, more sensitive than MRI and CUES. Non-ionizing, non-invasive and inexpensive in nature.

7. Presentation title 7 LITERATURE REVIEW Key takeaways from past researches. For detailed schedule, see management Note: Please make sure you are familiar with details of the schedule N otable gap in studies specifically focused on the classification of benign liver lesions such as FNH and HEM. Major studies focuses on Benign VS Malignant classification. Other studies utilized Ensemble of Neural Networks. Out of the many a few focused to include FNH and HEM into their classification along with other FLLs classification. SVM classifer was more efficient than other algorithms havin accu to Other studies utilized Ensemble of Neural Networks.

8. Presentation title 8 Dataset collection FNH HEM C entral scar with a well-defined hyperechoic area . P eripheral displacement of vessels. T hough central scar is not always present. large central artery is usually present with a spoke wheel like centrifugal flow and portal veins are absent. rather it is detected incidentally. bright, solid masses (hyperechoic) with clear, sharp borders (sharply defined). brightness is due to red blood cells in the tumour's vessels reflecting sound waves. <5 cm with round or oval shape 4 types echogenicity shown: Hyperechoic: Brighter than surrounding. H ypoechoic: Darker than surrounding. I soechoic: S imilar to surrounding. A nechoic: no echoes, complete dark.

9. 9 Sample images HEM (Hemangioma) FNH (Focal nodal Hyperplasia) Liver Dataset

10. Presentation title 10 METHODS OVERVIEW ROI extraction Size: 32 x 32 pixels (1,024 pixels) IROI (In-Region of Interest) feature vectors extracted from within the lesion. SROI (Surrounding Region of Interest) feature vectors extracted from outside the lesion.

11. 11 CAD system workflow Extraction technique: Gray Level Co-occurrence Matrix (GLCM). - Feature vectors: IROI (In-Region of Interest) feature vectors extracted from within the lesion. SROI (Surrounding Region of Interest) feature vectors extracted from outside the lesion. Ratio features calculated as the ratio of IROIs and SROIs. Feature Vector Construction Concatenation of 26 texture features: 13 texture features from IROI. 13 ratio features from SROI.

12. 12 M E T H O D O L O G Y Classification Classifier used: Support Vector Machine (SVM). Experimental setup: Different distances (d) and angles for GLCM. 3 types of feature vectors tested. Highest overall accuracy: 85% using concatenated features. - Individual accuracies: HEM: 93.3% FNH: 76.7%

13. Thank You

14. 14 Glcm FEATURES CALCULATED FOR RESEARCH Table : F1 to F13 GLCM Features Feature ID Feature Name 1 F 1 : Angular Second Moment (ASM) 2 F 2 : Contrast 3 F 3 : Correlation 4 F 4 : Sum of Squares-Variance 5 F 5 : Inverse Difference Moment (IDM) 6 F 6 : Sum Average 7 F 7 : Sum Variance 8 F 8 : Sum Entropy 9 F 9 : Entropy 10 F 10 : Difference Variance 11 F 11 : Difference Entropy 12 F 12 : Information Measures of Correlation-1 (IMC-1) 13 F 13 : Information Measures of Correlation-2 (IMC-2)