In recent years, computer-assisted medical imaging diagnosis has gradually become a research hotspot in this field. In order to better classify and identify medical image features, this study proposes an image recognition method that integrates convolutional neural networks and improved iterative deep learning based on adaptive learning. In the process, a randomized fusion improved convolutional neural network is introduced to cope with the multimodal feature extraction of medical images, and combined with improved iterative deep learning to avoid the loss of image data information, and finally complete the recognition of image information. The results show that the research method is experimented on the training set and the validation set. When the iteration is carried out to the 28th and 17th times, the system begins to stabilize, and the corresponding loss function values are 0.0124 and 0.0112 respectively. When the precision of the four algorithms is 0.900, the recall rates of the improved deep learning model, LeNet-5CNN model, IYolo-v5 model and the research method are 0.6232, 0.5791, 0.6774 and 0.8369 respectively. The recognition accuracy of the research method for the five diseases is significantly higher than 95%. The above results indicate that the research method has a fast convergence speed and accuracy, and can be widely used in image diagnosis and recognition of various types of diseases.