Click here to


Are you sure ?

Yes, do it No, cancel

Designing a Simple CNN Model in Terms of Size and Computational Complexity to Perform Classification Task On Medical Images

R Immanni1, G Valdes2*, Y Interian3, (1) University of San Francisco, San Francisco, CA, (2) University of California San Francisco, San Francisco, CA, (3) USF, San Francisco, CA


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Medical image datasets are fundamentally different from natural image datasets in terms of the number of training observations and number of classes. We hypothesized that compared to architectures used for natural images, those needed for medical imaging can be simpler. Here, we propose smaller architectures and show how they perform similarly while significantly saving training time and memory.

Methods: Existing state of the art CNN models like ResNet and MobileNet are used as base-line architectures. Methods used by us to compress the model include 1) using different convolution blocks that include grouped convolutions and depthwise separable convolutions 2) Delaying the increase in the number of filters to avoid capturing redundant features 3) Use little regularization to achieve better accuracy(as the models got simpler). Popular medical datasets from open data sources are chosen to test the model. The model performance is measured using ROC_AUC scores. As a proxy to quantify computational complexity and training time we calculate the number of parameters(in Millions) and GMACs for the computations

Results: All the models are trained under similar conditions i.e fixed epochs and learning rates. When testing Mobilenet, we were able to reduce number of parameters from 2.33M to 1.66M and the number of GMACs from 0.34 to 0.26 while attaining a similar roc_auc score on validation dataset. When we experimented with ResNet18, we were able to reduce the number of parameters from 11.18M to 554K(20 times), while increasing the performance from 0.804 to 0.858. Additionally, the number of GMACs were reduced from 1.91 to 0.22.

Conclusion: Although complex models are used to analyze Natural Images, simpler models can be used to train typical medical images without sacrificing accuracy. In this work, we have validated different techniques to reduce the complexity of models and as such reduce their complexity and training time


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email