Brain tumours can cause neurological symptoms and impair cognitive functions, depending on their location and size. Early detection allows for interventions that can prevent or minimize neurological deficits. By detecting tumours before they cause significant damage, healthcare professionals can employ treatment strategies to preserve brain function and improve the patient's quality of life. A decade back, many researchers implemented image segmentation approaches for tumour detection. These models need more Generalization because Segmentation models trained on specific datasets may need help to generalize well to different datasets or populations. Variations in imaging protocols, equipment, and patient populations can affect the performance of segmentation algorithms. Therefore, models trained on one dataset may not perform optimally on new, unseen datasets, requiring adaptation or retraining. Later, Neural Network models came into existence, and Convolutional Neural Networks (CNNs) have shown remarkable success in automating brain tumour detection. CNNs can learn hierarchical representations directly from the raw image data and automatically segment and classify brain tumours. They have been used for tumour segmentation, classification, and grading tasks. This paper studies various methodologies for tumour detection using CNN approaches. Among the studies conducted, transfer learning may not be the most efficient CNN approach for brain tumour detection; it's important to note that its limitations are in the context of complex and demanding tasks like tumour detection. Pre-trained architectures can still be effective for simpler image classification tasks or serve as a starting point for more advanced architectures through transfer learning or model adaptation.