Photothermal Therapy (PT) is an efficient scanning technique in biological research. However, rapid PT scanning and noise attenuation constitute a critical challenge in a clinical trial. To solve this problem, techniques based on Compression Sensors (CS) have recently been developed that use lower PT calculation costs and excellent accuracy. The majority of image results of prominence techniques depends on rare and inconsistent circumstances. It also takes considerable time to ensure that the photothermal information gathered empirically meets CS requirements. To address these constraints, a Hyperspectral Images Deep Learning (HSIDL-PT) and Maximize-Minimize Algorithm (MMA) method proposed in this paper to efficiently reduce under-represented artifacts in Neural Network (NN). The proposed method is to retrieve high photothermal images which used 5% of the total measuring information. Furthermore, the quantitative evaluation shows that the s HSIDL-PT technique improves performance by 30 percent when compared to the existingZ technique. with mean Structure Similarity Index (SSIM) = 0.974, Peak Signal-to-Noise Ratio (PSNR) = 29.88 dB and Standard Deviation of ± 0.007. Researchers propose MMA with Deep Learning, which would be the outcome of employing the limit for non-convex optimization, ethical and responsible that used the unifying MMA viewpoint; studies demonstrate that the proposed method outperforms the only one depending on linear maximizing.