Pneumonia has led to the rapid growth of mortality and morbidity which is now a global public health issue. This research provides the medical professional community specifically radiologists with an open-source model to continuously assist them and enhance their capabilities in reviewing complex chest X-ray images by identifying invisible patterns and characteristics of the pneumonia virus. A validation accuracy of 91%, 86%, and 89% was recorded for the three pieces of training while an accuracy of 86%, 84%, and 82% respectively. Taking the average of both resulted in an average validation accuracy of 88.6% and an accuracy of 84.0%. Consequently, the result of this study is better than some of the previous research works cited. However, in the future, more robust experiments are still needed to improve the model’s accuracy. Additionally, more data will be acquired to prevent the issue of data imbalance.
-
The initialisation of a sequential model (stacked on top of each other) using Keras, CNN layer of a fixed filter size for all four Con2D layers is (3,3) starting with 32 filters was defined, this simply means that each layer uses a 3x3 kernel to scan the input (X-ray image), max pooling with a pool size of (2,2) which reduces the spatial dimensions of the output from the former layer followed by an activation function (relu) and a regularise (L2).
-
The dropout layer is applied next which sets a fraction of the input units to prevent overfitting during training. The above pattern Conv2D – MaxPoling2D – Dropout is repeated three times with different filters for the Conv2D layers as 64, 128, and 256 respectively.
-
A flattened layer is introduced to convert the 2D metrics into a 1D vector space and a new dropout layer is added with the rate of 0.5. Lastly, a dense layer with 1 unit and a new activation function (sigmoid) is introduced. The idea of introducing a sigmoid activation function at this stage is to ensure that the output is the probability given the input point belongs to the class of 1.
To access the full technical report and findings click me