Concerning the problems of the difference in the number of samples between classes and the diversity of image background caused by environmental factors in the screening method of adolescent scoliosis based on back images and deep learning， a scoliosis screening method was designed， including steps such as back image data enhancement， back region extraction， and scoliosis diagnosis. Firstly， an improved diffusion generation model based on double residual U-Net structure and Convolutional Self-Attention Mechanism （CSAM） was proposed to generate high-quality pseudo-samples for the minority class back images， so as to balance the class distribution. Secondly， a back region extraction model with multi-loss constraint balance was designed to identify and extract back region from the back image， so as to eliminate the influence of image background difference on the diagnosis model. Thirdly， based on the selective kernel feature extraction and Spatial Pyramid Pooling （SPP） technologies， a classification model was constructed to realize the early screening and severity diagnosis of scoliosis through the back region. Finally， by integrating the above methods， the computer software and mobile software were developed to facilitate the actual back image acquisition and scoliosis screening business. Experimental results show that on the self-made scoliosis dataset， the proposed method achieves 98.64% and 73.06% accuracy in scoliosis early screening and severity diagnosis tasks， respectively， which is 2.52 and 6.48 percentage points higher than that of ResNet101. It can be seen that the proposed method can complete the diagnosis of scoliosis conveniently and quickly， and has certain application scenarios in large-scale rapid screening of scoliosis.

Concerning the problem that when the state prediction error is directly used as the intrinsic curiosity reward， the reinforcement learning agent cannot effectively explore the environment in the task with low correlation between state novelty and reward， an Intrinsic Curiosity Module with Reward Prediction Error （RPE-ICM） was proposed. In RPE-ICM， the Reward Prediction Error Network （RPE-Network） model was used to learn and correct the state prediction error reward， and the output of the Reward Prediction Error （RPE） model was used as an intrinsic reward signal to balance over-exploration and under-exploration， so that the agent was able to explore the environment more effectively and use the reward to learn skills to achieve better learning effect. In different MuJoCo （Multi-Joint dynamics with Contact） environments， comparative experiments were conducted on RPE-ICM， Intrinsic Curiosity Module （ICM）， Random Network Distillation （RND） and traditional Deep Deterministic Strategy Gradient （DDPG） algorithm. The results show that compared with traditional DDPG， ICM-DDPG and RND-DDPG， the DDPG algorithm based on RPE-ICM has the average performance improved by 13.85%， 13.34% and 20.80% respectively in Hopper environment.