Learning Causal Alignment for Reliable Disease Diagnosis

Authors: Mingzhou Liu, Ching-Wen Lee, Xinwei Sun, Xueqing Yu, YU QIAO, Yizhou Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the effectiveness of our method on two medical diagnosis applications, showcasing faithful alignment to radiologists. Code is publicly available at https://github.com/lmz123321/Causal_alignment. [...] In this section, we evaluate our method on two medical diagnosis tasks: the benign/malignant classification of lung nodules and breast masses. [...] We repeat 3 different seeds to remove the effect of randomness. [...] Table 1: Comparison with baseline methods on LIDC-IDRI and CBIS-DDSM datasets. [...] Table 2: Ablation study on LIDC-IDRI and CBIS-DDSM datasets. [...] Figure 4: CAM visualization. Each row denotes different cases.
Researcher Affiliation Academia Mingzhou Liu1 Ching-Wen Lee1 Xinwei Sun 2 Xueqing Yu1 Yu Qiao 3 Yizhou Wang4,1,5,6,7 1 School of Computer Science, Peking University 2 School of Data Science, Fudan University 3 School of Automation and Intelligent Sensing, Shanghai Jiao Tong University 4 Center on Frontiers of Computing Studies, Peking University 5 Institute for Artificial Intelligence, Peking University 6 Nat l Eng. Research Center of Visual Technology, Peking University 7 State Key Lab. of General Artificial Intelligence, Peking University Corresponding authors EMAIL EMAIL, EMAIL, EMAIL
Pseudocode Yes Algorithm 1 Causal alignment training Input: Data D, Output: Decision model fθ, Hyperparameters: Sparsity regularization α, weight of alignment loss λ, learning rate η. 1: while not converged do 2: **Forward pass 3: Compute Lce. 4: Optimize (2) to obtain x and compute Lalign using (3). 5: Compute L Lce + λLalign. 6: **Back propagation 7: Estimate θLalign with conjugate gradient. 8: Update θ: θ θ η θL. // or Adam 9: end while
Open Source Code Yes Code is publicly available at https://github.com/lmz123321/Causal_alignment.
Open Datasets Yes We consider the LIDC-IDRI dataset Armato III et al. (2011) for lung nodule classification and the CBIS-DDSM dataset Lee et al. (2017) for breast mass classification.
Dataset Splits Yes We split the dataset into training (n = 731), validation (n = 238), and test (n = 244) sets. The CBIS-DDSM dataset ... We follow the official dataset split, with 691 masses in the training set and 200 masses in the test set.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No We use the Adam optimizer and set the learning rate as 0.001. We adopt the Torch Opt Ren et al. (2022) package to implement the conjugate gradient estimator.
Experiment Setup Yes We use the Adam optimizer and set the learning rate as 0.001. We parameterize the attributes prediction network fθ1 with a seven-layer Convolutional Neural Network (CNN), and train it for 100 epochs with a batch size of 128 for each iteration. For the classification network fθ2, we parameterize it with a two-layer Multi-Layer Perceptron (MLP), and train it for 30 epochs with a batch size of 128. Please refer to Appx. B for details of the network architectures. For the hyperparameters α1 in (7) and α2 in (6), we set them to α1 = 0.01, α2 = 0.0005 for LIDC-IDRI and α1 = 0.07, α2 = 0.0005 for CBIS-DDSM, respectively. For both datasets, we set λ1 = λ2 = 1 in (5).