Scale-Space Hypernetworks for Efficient Biomedical Image Analysis
Authors: Jose Javier Gonzalez Ortiz, John Guttag, Adrian Dalca
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our approach using several medical image segmentation and registration tasks including different biomedical domains and image types. We demonstrate that the proposed hypernetwork based approach is able to learn models for a wide range of feature rescaling factors, and that inference networks derived from the hypernetwork perform at least as well, and in most cases better than, networks trained with fixed rescaling factors. SSHNs enable us to find that a wide range of rescaling factors achieve similar accuracy results despite having substantially different inference costs. |
| Researcher Affiliation | Academia | Jose Javier Gonzalez Ortiz MIT CSAIL Cambridge, MA josejg@mit.edu John Guttag MIT CSAIL Cambridge, MA guttag@mit.edu Adrian V. Dalca MIT CSAIL & MGH, HMS Cambridge, MA adalca@mit.edu |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | We provide additional platform and reproducibility details in the supplement, and we will publicly release our code. |
| Open Datasets | Yes | We use four popular and quite different biomedical imaging datasets OASIS [25, 50], Pan Dental [1], WBC [76] and CAMUS [43]. |
| Dataset Splits | Yes | In each dataset, we trained using a (64%, 16%, 20%) train, validation, test split of the data. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory amounts used for running the experiments. |
| Software Dependencies | No | The paper mentions software components like 'Adam optimizer', 'U-Net architecture', and 'Leaky ReLU activations', but does not provide specific version numbers for any software libraries or frameworks, which are necessary for full reproducibility. |
| Experiment Setup | Yes | For our experiments, we use the U-Net architecture, since it is the most widely-used segmentation CNN architecture in the medical imaging literature [30, 58]. We use bilinear interpolation layers to downsample or upsample a variable amount based on the scale factor ϕ for downsampling. For the downsampling layers we multiply the input dimensions by ϕ and round to the nearest integer. For the upsampling layers we upsample to match the size of the feature map that is going to be concatenated with the skip connection. We use five encoder stages and four decoder stages, with two convolutional operations per stage and Leaky Re LU activations [48]. For networks trained on OASIS Brains we use 32 features at each convolutional layer. ... We implement the hypernetwork using three fully connected layers. ...The hidden layers have 10 and 100 neurons respectively and Leaky Re LU activations [48] on all but the last layer, which has linear outputs. Hypernetwork weights are initialized using Kaiming initialization [23], and bias vectors are initialized to zero. ... We use the Adam optimizer [38] and we train networks until the loss in the validation set stops improving. During training we sample the rescaling from U(0, 0.5) |