Alzheimer’s Disease Early Detection Using a Densenet-121
13
is bad because other classes are classified as AD, but in a context of pattern
recognition that always has risks and costs, it is in favor because it is pessimistic
and in medicine that can reduce risk and future costs. In the same manner, the
bad number of LMCI is also acceptable because the class is mostly classified
as AD. Therefore, considering the economic restrictions, the final figures of 84%
accuracy, 84% specificity (micro) and 81% sensitivity (micro) are acceptable. We
chose to report final micro-average figures instead of macro-average because in a
multi-class classification setup, micro-average is preferable when there is a class
imbalance. However, as it can be noticed the macro average and the weighted
average are better.
6 Conclusions and Future Work
The use of free-of-charge resources limited this study. With this restriction, we
explored a low-cost way to generate a deep artificial neural network that shows
good performance metrics. We demonstrate that the model can still be improved.
This prediction model can be useful in developing countries if user interface and
interpretation are added and it has the potential of being used in remote medicine
contexts.
In the future, we want to create a user interface for the diagnosis of AD. We
can do this based on the implementation of Chester [5], a computerized chest
X-ray disease prediction system that is delivered on the web. With the recent
creation of tools such as ONNX and TensorFlow.js, PyTorch-trained models
can be converted to work in the browser and compute using WebGL [5]. This
interface would have not only prediction but also interpretation or explanation
through relevance maps or heat maps.
Last, to contribute to reproducibility and transparency in academic work, we
provide the source code of our DNN at https://github.com/bsolano/AlzheimerResNets.
References
1. Alzheimer’s Disease Neuroimaging Initiative: Study Design (2017). http://adni.
loni.usc.edu/study-design/
2. B¨ ackstr¨ om, K., Nazari, M., Gu, I.Y., Jakola, A.S.: An efficient 3D deep convolutional network for Alzheimer’s disease diagnosis using MR images. In: 2018 IEEE
15th International Symposium on Biomedical Imaging (ISBI 2018), pp. 149–153,
April 2018. https://doi.org/10.1109/ISBI.2018.8363543
3. Carneiro, T., Medeiros Da N´ oBrega, R.V., Nepomuceno, T., Bian, G., De Albuquerque, V.H.C., Filho, P.P.R.: Performance analysis of Google colaboratory as
a tool for accelerating deep learning applications. IEEE Access 6, 61677–61685
(2018). https://doi.org/10.1109/ACCESS.2018.2874767
4. Cheng, D., Liu, M.: CNNs based multi-modality classification for AD diagnosis.
In: 2017 10th International Congress on Image and Signal Processing, BioMedical
Engineering and Informatics (CISP-BMEI), pp. 1–5, October 2017. https://doi.
org/10.1109/CISP-BMEI.2017.8302281
13
is bad because other classes are classified as AD, but in a context of pattern
recognition that always has risks and costs, it is in favor because it is pessimistic
and in medicine that can reduce risk and future costs. In the same manner, the
bad number of LMCI is also acceptable because the class is mostly classified
as AD. Therefore, considering the economic restrictions, the final figures of 84%
accuracy, 84% specificity (micro) and 81% sensitivity (micro) are acceptable. We
chose to report final micro-average figures instead of macro-average because in a
multi-class classification setup, micro-average is preferable when there is a class
imbalance. However, as it can be noticed the macro average and the weighted
average are better.
6 Conclusions and Future Work
The use of free-of-charge resources limited this study. With this restriction, we
explored a low-cost way to generate a deep artificial neural network that shows
good performance metrics. We demonstrate that the model can still be improved.
This prediction model can be useful in developing countries if user interface and
interpretation are added and it has the potential of being used in remote medicine
contexts.
In the future, we want to create a user interface for the diagnosis of AD. We
can do this based on the implementation of Chester [5], a computerized chest
X-ray disease prediction system that is delivered on the web. With the recent
creation of tools such as ONNX and TensorFlow.js, PyTorch-trained models
can be converted to work in the browser and compute using WebGL [5]. This
interface would have not only prediction but also interpretation or explanation
through relevance maps or heat maps.
Last, to contribute to reproducibility and transparency in academic work, we
provide the source code of our DNN at https://github.com/bsolano/AlzheimerResNets.
References
1. Alzheimer’s Disease Neuroimaging Initiative: Study Design (2017). http://adni.
loni.usc.edu/study-design/
2. B¨ ackstr¨ om, K., Nazari, M., Gu, I.Y., Jakola, A.S.: An efficient 3D deep convolutional network for Alzheimer’s disease diagnosis using MR images. In: 2018 IEEE
15th International Symposium on Biomedical Imaging (ISBI 2018), pp. 149–153,
April 2018. https://doi.org/10.1109/ISBI.2018.8363543
3. Carneiro, T., Medeiros Da N´ oBrega, R.V., Nepomuceno, T., Bian, G., De Albuquerque, V.H.C., Filho, P.P.R.: Performance analysis of Google colaboratory as
a tool for accelerating deep learning applications. IEEE Access 6, 61677–61685
(2018). https://doi.org/10.1109/ACCESS.2018.2874767
4. Cheng, D., Liu, M.: CNNs based multi-modality classification for AD diagnosis.
In: 2017 10th International Congress on Image and Signal Processing, BioMedical
Engineering and Informatics (CISP-BMEI), pp. 1–5, October 2017. https://doi.
org/10.1109/CISP-BMEI.2017.8302281