All-in-one CT generation from DICOM using hybrid multi-window segmentation, TransU-Net, and Pix2Pix

Computed tomography (CT) interpretation often requires switching among soft-tissue, lung, and bone window settings, increasing reading time and cognitive burden. Few studies have investigated the direct generation of a clinically usable all-in-one (AIO) CT image from routine DICOM data while preserving tissue-specific diagnostic information. Since AIO images are secondary representations synthesized from multiple window settings rather than directly acquired data, constructing a gold-standard reference is both labor-intensive and subjective. This study aimed to develop a silver-standard AIO reference and a deep learning framework for direct AIO CT synthesis from routine CT data. Reference AIO images were constructed using a hybrid pipeline combining conventional image processing and deep segmentation, followed by fusion of bone, lung, and soft-tissue regions. A Pix2Pix-based conditional generative adversarial network was then trained to synthesize AIO CT images end-to-end. Segmentation performance was evaluated on CT-ORG, and synthesis performance was assessed using the structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), and Fréchet inception distance (FID). Clinical usefulness was further examined through expert review and a preliminary lesion detection study using 64 breast cancer follow-up CT examinations. On CT-ORG, TransU-Net achieved the best segmentation performance, with a mean Dice similarity coefficient of 0.934. For AIO image synthesis, Pix2Pix outperformed CycleGAN and AttentionGAN, achieving an SSIM of 0.99, a PSNR of \(43.42 \pm 1.39\) dB, and an FID of \(11.03 \pm 8.64\) . Expert review demonstrated good overall image quality and favorable depiction of abdominal, lung, and bone structures. In the preliminary clinical feasibility assessment, lesion detection accuracy was 85.94% for lung lesions and 87.50% for bone lesions. The proposed framework enables generation of clinically informative AIO CT images from standard DICOM data and may support more efficient multi-window CT interpretation in clinical practice.