Identifying lung cancer can be done effectively by examining CT scan images. It is necessary to have an intelligent diagnostic system. The recent evolution of image categorization systems has been influenced by the use of deep convolution neural networks (CNN). In this research work, a new hybrid paradigm with images combines a modified deep transfer learning EfficientNet and a masked autoencoder for distribution estimation (MADE). By using MADE before classification in lung cancer classification can facilitate feature learning, dimensionality reduction, uncertainty estimation, handling imbalanced data, transfer learning, and model interpretability, ultimately leading to improved classification performance and better utilization of available data. The proposed model (Mask-EffNet) works in two phases. In the initial phase feature extraction is done by using MADE and the classification of different types is carried out in the subsequent phase using a pre-trained EfficientNet model. Mask-EffNet is tested using EfficientNetB7 variation. The research is conducted on the "IQ-OTH/NCCD" benchmark dataset, which consists of lung cancer patients classed as benign, malignant, or normal depending on whether they have lung cancer or not. The model Mask-EffNet attained an accuracy of 98.98% and a ROC score ranging from 0.9782 to 0.9872 on the test set. We examined the effectiveness for suggested pre-trained Mask-EffNet to that of various additional pre-trained CNN designs. The anticipated results show that the Mask-EffNet based on EfficientNetB7 prevails over different CNNs with regard of both accuracy as well as effectiveness.

There is a level of saturation obtained about Online or Printed character recognition about Indian script. Comparing the research of different Indian languages obtained yet a few researches has been done for Odia characters. There are still more research remains valid while recognizing handwritten Odia script. To overcome this challenges a proper dataset is required for experimentation. The handwritten data collected from various Odia writers. There are four process carried out to generate dataset i.e. data collection, Pre-processing, Segmentation and classification. Data is collected from different sources like school and college students, village people etc. arranged them into different age groups, gender and education level. In pre-processing selection of image, noise removal, normalization, conversation of grey scale to binary image, then converted binary images to inverted images. In this paper we have focused only handwritten simple characters that included vowels, consonants, numbers and special characters. We have created 95000 of character data of different varieties compared with other availability dataset and to find its better accuracy we have used `deep learning approach like InceptionV3 with different batch size and epoch level and find its accuracy. We obtain an accuracy of 84.50 for the used model for our collected dataset. Our handwritten data is available publically and it will be useful for the researcher to continue their future research work and for the time being it can be available on request.