In this paper, automated biomedical waste detection and classification using various convolutional neural network models are explored for comparison and analysis of model accuracy and model loss obtained from implemented convolutional neural network (CNN) models. Specifically, the topic discussed here includes the architectures that utilise main layer configuration and activation functions for an architecture used in an analysis model. Here, the comparison and analysis of all models are performed based on their architectures; some of the models have larger architecture such as ResNet101, allowing them to be trained on datasets with more complex applications that can capture more complex features and utilises residual connections help mitigate vanishing gradient problem. GoogleNet with inception modules efficiently computes and has better feature extraction, specifically helpful for catching variations in biomedical waste images. Bio-medical waste dataset with 2572 images contains sharps, infectious, and pathological waste. After hyperparameter fine-tuning of some of the mentioned hyperparameters, such as learning rate, batch size, and dropout rate, from the abovementioned auto-tuning methods, the loss rate dramatically reduced while showing improvement in accuracy. This is crucial for the models’ generalisation capability. It is proved that it has models that are trained with systematic hyperparameter optimisation as the objective of the research. Describing preprocessing techniques, such as normalisation and image resizing, the paper gives a robust base in using them as essential input standardisation methods so that any level of consistency can be reached with model performance.