Conference Papers
Permanent URI for this collectionhttps://idr.nitk.ac.in/handle/123456789/28506
Browse
Search Results
Item Osteosarcoma Bone Cancer Detection(Springer Science and Business Media Deutschland GmbH, 2025) Payani, C.A.; Gupta, C.; Vamsidhar, K.; Bhat, P.; Patil, N.Osteosarcoma is a type of bone cancer commonly found in the elongated bones found in the upper and lower limbs. The precise cause is unknown, but experts believe it’s linked to changes in the DNA of the bones, resulting in the growth of abnormal and harmful bone tissue. If caught early, osteosarcoma is treatable, with about 75% people cured when the cancer hasn’t spread to other body parts. Analyzing biopsy samples can be time-consuming, but there are advanced computer programs, known as supervised deep learning methods, that can help speed up the process and enhance the efficiency of the diagnosis. Previous studies have already evaluated the performance of deep learning models such as VGG16, VGG19, DenseNet201, and ResNet101, among which ResNet101 performed better with 90.36% accuracy. When it comes to understanding complex image features, previous models lack behind. We propose EfficientNetV2, Xception, and InceptionV3 models, among which Xception outperformed other models with 94.5% accuracy on the image dataset. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.Item Exploring Various Data Mining Techniques to Predict Heart Disease(Springer Science and Business Media Deutschland GmbH, 2025) Makam, S.K.; Hiranmayi, M.Y.; Kumar, P.; Bhat, P.; Patil, N.One of the main causes of fatalities in the global population is cardiovascular disease (CVD), commonly called heart disease. Early detection of CVD risks is a major area of interest in clinical data analysis. This study focuses on devising strategies for improving the predictive abilities of CVD risk detection algorithms. We experiment with binary and multiclass classification techniques on public UCI machine learning repository datasets, namely, Cleveland for training and Statlog and Hungarian for evaluation. The techniques include feature selection by best subset generation and data balancing using Binary and Multiclass SMOTE and their variants. Every technique is assessed by tenfold cross-validation on six classifiers: K-Nearest Neighbors (KNNs), Naive Bayes, Logistic Regression (LR), Support Vector Machine (SVM), Neural Network, and Vote (a hybrid technique combining Naïve Bayes and Logistic Regression). Experimental results show a rise in average classifier F1-score of 4.36% after feature selection and Binary SMOTE. Top-performing models include Logistic Regression, Neural Networks, and Voting. KNN shows a significant rise of 8.5 and 5.05% in accuracy, after employing Binary and Multiclass SMOTE techniques, respectively. Multiclass SMOTE results can be used as a benchmark but possess scope for further research and enhancement. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.Item Automated Colorization of Grayscale Images Using Superpixels and K-Means Clustering(Springer Science and Business Media Deutschland GmbH, 2025) Kulkarni, B.C.; Teja, B.; Hegde, A.R.; Bhat, P.; Patil, N.The process of transforming grayscale photos into aesthetically pleasing color images is called colorization. Convincing the audience of the realism of the outcome is the primary objective of colorization. Natural scenery makes up the majority of the grayscale photographs that require colorization. A broad range of colorization techniques have been created over the past 20 years; these vary from algorithmically basic procedures that need time and energy due to inevitable human participation to more complex ways that are also more automated. The complex field of automatic conversion mixes deep learning, machine learning, and art. Most of the earlier works which use deep learning, use every pixel values to train the models which is computationally expensive. We present a methodology for colorizing grayscale images using convolutional neural network (CNN), our method uses a combination of superpixel segmentation and K-Means clustering to significantly reduce number of pixel values. The process begins with the conversion of grayscale images to superpixels, which are perceptually uniform regions that aid in efficient colorization. Subsequently, K-Means clustering is applied within each superpixel to identify dominant color clusters, followed by quantization of color information to simplify representation. The prepared input, comprising grayscale images and quantized color information, is then fed into a CNN for colorization, leveraging spatial coherence and semantic context to predict plausible colors for grayscale pixels. The proposed methodology is evaluated on a diverse set of grayscale images, demonstrating its effectiveness in producing vibrant and visually appealing colorized outputs. Through experiments and analysis, we showcase the potential applications and benefits of the proposed approach in historical photograph restoration, movie colorization, and other domains requiring accurate and efficient grayscale image colorization. We use SSIM and PSNR as our evaluation metrics. SSIM is calculated based on the similarity of the luminance and brightness values of the considered and obtained rgb images for the grayscale images, and PSNR is calculated using Mean Squared Error (MSE) of the peak signal values within images. Our methodology’s SSIM and PSNR for the considered flower class is 81.5 and 25.6, respectively. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.Item A Hybrid Weighted Loss Function for Enhanced Protein Interaction Site Prediction(Springer Science and Business Media Deutschland GmbH, 2025) Bhat, P.; Patil, N.Accurately predicting protein interaction sites is crucial for applications such as protein design, drug discovery, and functional protein analysis. However, a significant challenge in this task arises from the inherent class imbalance between interacting and non-interacting sites in protein datasets. While data augmentation techniques are commonly used to mitigate this imbalance, they often introduce noise, potentially reducing prediction accuracy. In this study, we present a novel approach to improve protein interaction site prediction by developing a customized loss function that combines focal loss and cost-sensitive loss, specifically designed to address class imbalance without relying on data augmentation. Our model, which integrates graph convolutional networks (GCNs) to process evolutionary and structural features of proteins, is evaluated using robust performance metrics suited for imbalanced data: Matthews Correlation Coefficient (MCC) and Area Under Precision-Recall Curve (AUPRC). We evaluate the proposed method on the Test_60 dataset, achieving an MCC of 0.342 and an AUPRC of 0.425, providing a modest improvement over the standard cross-entropy loss function. These findings highlight the effectiveness of our tailored loss function in handling class imbalance and improving prediction performance in protein interaction site prediction. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.