1
$\begingroup$

I'm currently working on a project to classify the medical images as normal and abnormal using convolution neural network with input as feature data (Statistical feature values extracted from medical images). Everywhere i see input as an image resized to the size of the network and hence can anyone guide with this and that'll be a great help. Thanks.

$\endgroup$
  • 1
    $\begingroup$ Do your extracted features keep a notion of spatiality? If not, convolutional filters are likely not what you want, a fully connected net would be a preferable aproach. $\endgroup$ – Jenkar Oct 20 '17 at 9:46
  • 1
    $\begingroup$ There are features that represent the voxel of an image (Spatial 3D representation of an pixel) and also few features that define shape features. So can you give me a clear idea about what is spatiality in a feature value. Thanks. $\endgroup$ – Akhil B.G. Oct 20 '17 at 11:15
1
$\begingroup$

If you have doubts regarding the image resize reasons and solutions of network sizes, a great paper [1] is describing the solution of several authors through it. A detailed review is done in this paper regarding the Deep Learning (DL) techniques and applicability in the field of Medical Imaging (MI) analysis.

The authors used a Convolutional Neuro Network (CNN) with Max Pooling Layers and Fully Connected Layers. The CNN architecture for classification of the medical images has several classes accepting a path of 32 x 32 from an original 2D medical image. From here, each CNN Layer generates a feature map from different sizes and the Pooling Layers reduce the feature map sizes to be possible to transfer between the following layers.

Another interesting approach is the one presented on the topic of Pedestrian Detection [2]. Here, the original input size is downscaled from 224 x 224 x 3 to 64 x 64 x 3, in order to ease the computational-expense associated with the CNN. As a result, the authors adjust the Fully Connected Layers to the hereby modification, by randomly initializing and resizing them.

You have several other projects and articles explaining their efforts to improve computability. One new hype in relation to your problem is the workaround breast cancer diagnosis [3] and Radiomics [5]. Since the active breast lesion detection with medical annotations [4] (i.e., segmentation and labelling) is of chief importance to be little time-consuming. I recommend following the BreastScreening project, a project that joins contributions from both MIMBCD-UI and MIDA projects. The BreastScreening project aims to develop several AI-Assisted techniques to improve the medical diagnostic workflow by automatizing several parts of it. The goal is to develop an Assistant promoting a higher Radiology Room (RR) performance and bringing higher health care results for the patient. The new paradigms of MI in the Health Informatics (HI) field are many and it just started. Good to know to have more and more people start working in this field of research.

[1] Anwar, Syed Muhammad, Muhammad Majid, Adnan Qayyum, Muhammad Awais, Majdi Alnowami, and Muhammad Khurram Khan. "Medical Image Analysis using Convolutional Neural Networks: A Review." Journal of medical systems 42, no. 11 (2018): 226.

[2] Ribeiro, David, Gustavo Carneiro, Jacinto C. Nascimento, and Alexandre Bernardino. "Multi-channel Convolutional Neural Network Ensemble for Pedestrian Detection." In Iberian Conference on Pattern Recognition and Image Analysis, pp. 122-130. Springer, Cham, 2017.

[3] Maicas, Gabriel, Gustavo Carneiro, Andrew P. Bradley, Jacinto C. Nascimento, and Ian Reid. "Deep reinforcement learning for active breast lesion detection from dce-mri." In International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 665-673. Springer, Cham, 2017.

[4] Calisto, Francisco M., Alfredo Ferreira, Jacinto C. Nascimento, and Daniel Gonçalves. "Towards Touch-Based Medical Image Diagnosis Annotation." In Proceedings of the 2017 ACM International Conference on Interactive Surfaces and Spaces, pp. 390-395. ACM, 2017.

[5] Gillies, Robert J., Paul E. Kinahan, and Hedvig Hricak. "Radiomics: images are more than pictures, they are data." Radiology 278, no. 2 (2015): 563-577.

| cite | improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.