DEVELOPING A DIAGNOSTIC MODEL FOR EARLY DETECTION OF PULMONARY DISEASES USING CT SCAN IMAGING

Abstract

Rapid and appropriate distinction between benign and malignant lung nodules on CT is essential for management time gathering, unnecessary morbidity minimization as well as mortality. Manual reads are laborious and subject to intra-reader variability, and many health systems do not have the rccssary radiology resources, particularly in developing-world settings or for resource-limited locations. We investigate a hybrid pipeline where we connect CNN backbones frozen with the ImageNet weights followed by a linear SVM classifier after preprocessing is applied (z-score normalization and PCA [95% variance accounted for]). CT slices are resized to 224×224, normalized [0;1], and jittered (rotated, shifted, stretched,zoomed,and horizontally flipped); optional unsharp masking is used to highlight edges of lesions. We use ResNet, VGG16, DenseNet121, InceptionV3 and EfficientNetB0 as feature extractor (include_top=False) and rely on a GAP or light weighting head to get the embeddings; SVM output scores are calibrated with probability-calibrated SVM outputs using Platt learing. Evaluation is performed on a balanced test set (N = 300; 150 benign / 150 malignant) with accuracy, macro- and weighted-average precision/recall/F1 and class-wise results. The highest scores were achieved by EfficientNetB0–SVM (Accuracy = 0.98; Macro/Weighted Precision = 0.98; Recall = 0.98; F1-score = 0.98), followed by InceptionV3–SVM (0,97) and VGG16–SVM (0,96). DenseNet121–SVM and ResNet121–SVM achieved 0.93 and 0.92, respectively. Class-wise analyses show both EfficientNetB0–SVM and InceptionV3–SVM reached balanced high precision and sensitivity for both classes (Benign F1 = 0.98, Malignant F1 = 0.98), although the former just slightly favored benign sensitivity (Malignant Recall = 0.97).