Abstract:
Lung cancer is one of the type malignant cancer. It occurs in a human body as a result of
the uncontrolled growth of cells in lung tissues. Upon diagnoses the death rate of lung
cancer patients is high as compared to survival rate. The symptoms of lung cancer
manifest only in the advanced stages, therefore it’s extremely difficult to detect in its
initial stage. Computed tomography (CT) scans have made lung cancer detection possible
in its initial stage which is helpful for both diagnosis as well as its treatment. Computer
aided detection (CAD) systems are functional for malignant nodule detection. A CAD
system for lung cancer classification uses pre-processed 3D CT scan image slices given
to 3D-CNN with rectified linear unit (ReLU) to classify malignant nodules. ReLU
function does not trigger all the neurons at the same time which makes it proficient and
easy for approximation. Although ReLU falls on the negative side of the graph, the
weights are not updated during backward propagation because the slope is zero. Due to
this fact, dead neurons are generated, which are never activated. In addition, currently no
particular 3D CNN algorithm is available that can provide a generic classification of lung
cancer nodule detection. Existing CAD models and state-of-the-art methods are available
but they have lack of configurability. In this work, we have performed comprehensive
analysis and have discussed each step in detail in order to develop a configurable, efficient
method that can work with generic dataset with minimum pre-processing. Our proposed
method allows users to select between pre-processing, feature extraction and activation
function algorithms and help find the configuration which results in minimum error for a
particular dataset to achieve high accuracy and optimistic solution.
