TCAM-Diff: Triplane-Aware Cross-Attention Medical Diffusion Model

Authors: Zhenkai Zhang, Krista A. Ehinger, Tom Drummond

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on three different scales of medical datasets, Brain Tumour 128 128 128, Pancreas 256 256 256, and Colon 512 512 512, demonstrate outstanding results. We utilized MSE and SSIM to assess reconstruction quality and leveraged the Wasserstein Generative Adversarial Network (W-GAN) critic to assess generative quality. Comparisons with existing approaches show that our method gives better reconstruction and generation results than other encoder-decoder methods with similar sized latent spaces.
Researcher Affiliation Academia School of Computing and Information Systems, The University of Melbourne EMAIL, EMAIL
Pseudocode No The paper describes the model architecture and methodology but does not include any explicit pseudocode or algorithm blocks. Figure 1 provides an overview of the architecture, not an algorithmic procedure.
Open Source Code No The paper does not contain an explicit statement about releasing the source code for the described methodology or a link to a code repository. It only mentions using publicly available code for baseline models.
Open Datasets Yes We conduct the experiments on three publicly available datasets with varying scales: the Bra TS dataset includes 750 4D MRI volumes (484 training, 266 testing)... (Menze et al. 2014b; Simpson et al. 2019). The Pancreas Tumour dataset contains 420 3D CT volumes (282 training set, and 139 testing set) with varying dimensions (Simpson et al. 2019), The Colon Cancer dataset includes 190 3D CT volumes with 126 training and 64 testings, with diverse shapes of each scan (Simpson et al. 2019).
Dataset Splits Yes the Bra TS dataset includes 750 4D MRI volumes (484 training, 266 testing)... The Pancreas Tumour dataset contains 420 3D CT volumes (282 training set, and 139 testing set)... The Colon Cancer dataset includes 190 3D CT volumes with 126 training and 64 testings...
Hardware Specification Yes Our experiments were conducted on A100 GPUs, each with 80GB of GPU RAM.
Software Dependencies No The paper mentions using the MONAI toolkit (The MONAI Consortium 2020) and an Adam optimizer, but it does not specify version numbers for these or other software dependencies like Python, PyTorch, or CUDA.
Experiment Setup Yes The hyperparameters in our experiments included an Adam optimizer with a 3e 5 learning rate for the Bra TS dataset and a 1e 5 learning rate for higher-resolution datasets. The loss function weights were set to λ1 = 1e 2 and λ2 = 1e 3 for Bra TS, and λ1 = 0 for higher-resolution datasets. To ensure stability, we applied gradient clipping with a max norm of 1.0.