Thyroid carcinoma grows in the thyroid gland, a butterfly-shaped gland at the base of the neck. Thyroid carcinoma is one of the most common endocrine malignancies when compared to other carcinomas. Researchers gain an understanding of methods to categorise slides using whole slide images (WSI) taken by using e-scanners in the clinics (benign or malignant). However, in a thyroid WSI, the critical section that supports the detection result may be smaller, and only the worldwide label can be collected, making straight use of supervised learning framework impractical. Furthermore, because of clinical detection of thyroid cells necessitates the utilisation of various visual features in variety of scales, a traditional method of extraction of features may not provide adequate results. We propose a poorly guided multi-instance-based learning approach for thyroid cytopathological detection was based on learning algorithm with Multi-Scale Feature Fusion (MSF) utilising Convolutional Neural Network (CNN) in this research. We approach every WSI as a pouch, with several instances corresponding to the various parts of a WSI. The architecture has been taught to automatically identify the main areas and classify them. We also suggest a feature fusion framework in which minimal features are merged in the resulting feature map with instance-level awareness model is included, resulting in improved classification results.The proposed model, which is firstly trained and validated on acquired clinical data, achieves a 93.2 percent accuracy rate, outperforming all other techniques. We also put our model to the test on a publicly available histopathology dataset, and it outperformed in ultra-modern of deep multi-instance technique. A web application was developed to collect user data and accurately predict the type of disease.