Contributed Papers

Title: Classification of Pneumothorax in Chest X-ray images using Deep Neural Network

Authors: Manjeet Kaur, Ankita Pandey, and Arun Kumar from G. B. Pant University of Agriculture and Technology.

Abstract: Currently, lung diseases are extremely common throughout the globe and a few of which include chronic obstructive pulmonary disease, pneumonia, pneumothorax, tuberculosis. Air in the pleural cavity is thought to be “pneumothorax". This is a state when the lung surface or chest wall is breached, allowing air to enter the pleural space and causing the lung to collapse, it is a serious situation and might be life-threatening. In this study, we have taken chest X-ray images and then used the Deep Transfer Learning technique along with Machine Learning to detect the presence or absence of pneumothorax in chest X-ray images. We have aggregated the unique and computational attributes of Deep Learning and Machine Learning. To extract image features, the deep Transfer Learning based pre-trained Residual Network is used. For binary classification of pneumothorax, Support Vector Machine is employed to generate the optimal decision boundary (hyperplane). To achieve effective outcomes, a balancing data technique with augmentation is used to create a balance between the training and validation dataset, as well as an automatic adjusting learning rate technique called "ReduceLROnPlateau" to monitor validation loss and obtain optimal learning rate. This research work has set a new record with a good performance by achieving state-of-the-art results as 0.8831 in terms of AUC and 0.4375 in terms of loss. The Precision, Recall, and the f1 score obtained as 0.9285, 0.8125, and 0.867, respectively.

​

​

​

Title: Opinion Mining of Japanese Product Reviews using Translators, Word Embedding and ML Techniques

Authors: Ramanna T, P Deepa Shenoy, and Venugopal K R, UVCE, Bangalore University.

Abstract: Most individuals rely on translation to communicate and comprehend other languages. The main driving force behind this study is determining which translator, along with the blend of word embedding and Machine Learning(ML)techniques, when working with the Japanese language, has greater accuracy. Business owners struggle to understand how their products perform in Japan and receive feedback to improve it. Japanese is among the most complex languages to read and write. The issue can be resolved by classifying reviews and ratings into separate groups, such as excellent and negative so that business owners can improve their products based on the reviews. This paper analyzes the Amazon Product Review dataset in the Japanese language, which contains reviews and ratings. These reviews are translated into English using three different translators(Google, Microsoft, and Python). Then combination of word embedding(count vector and TF-IDF) and ML techniques(K Nearest Neighbor(KNN), Support Vector Machine(SVM), and Logistic Regression(LR)) is applied to the data for each translator. The outcome of this study is that the Python Translator along with the blend of TF-IDF and SVM outperforms all the translators with an accuracy of 80%.

​

​

​

Title: MDLCSL: A Multilayer Deep Learning Model for Classification of Sweet Lime Leaves

Authors: Divyansh Thakur ( IIIT Una ), Jaspal Saini ( IIIT Una ), and Srikant Srinivasan (Plaksha University)

Abstract: Artificial Intelligence (AI) has a crucial impact on human life by using automation to ease workload. Many fields show AI's abilities and it's being utilized to accomplish intricate tasks efficiently and effectively. This work explores the use of AI, specifically Deep Learning (DL), in agriculture through the classification of sweet lime leaves. A custom dataset of 4000 images was created and manually classified into four classes (Healthy, Infected, Decayed, Other). A multilayer DL model was developed using Convolutional Neural Networks (CNN) for leaf classification. The model was evaluated using five performance parameters: accuracy, loss, F1-Score, recall, and precision, achieving an accuracy of 99.05% with a low loss of 0.003, F1-Score and recall of 0.96, and precision of 0.98. The proposed model was benchmarked against InceptionV3 and VGG16 and demonstrated its superiority.