[Paper Review] A Deep Convolutional Neural Network for Lung Cancer Diagnostic

Mehdi Fatan Serj, Bahram Lavi+2 more|arXiv (Cornell University)|Apr 22, 2018

Lung Cancer Diagnosis and Treatment6 references42 citations

TL;DR

The paper introduces a deep CNN architecture (dCNN) for binary lung cancer classification from CT images and reports strong performance on Kaggle Data Science Bowl 2017, with high sensitivity and specificity and competitive log-loss.

ABSTRACT

In this paper, we examine the strength of deep learning technique for diagnosing lung cancer on medical image analysis problem. Convolutional neural networks (CNNs) models become popular among the pattern recognition and computer vision research area because of their promising outcome on generating high-level image representations. We propose a new deep learning architecture for learning high-level image representation to achieve high classification accuracy with low variance in medical image binary classification tasks. We aim to learn discriminant compact features at beginning of our deep convolutional neural network. We evaluate our model on Kaggle Data Science Bowl 2017 (KDSB17) data set, and compare it with some related works proposed in the Kaggle competition.

Motivation & Objective

• Motivate early, accurate lung cancer detection using deep learning on medical CT images.
• Develop a discriminative, compact feature learning CNN architecture for binary classification (cancer vs. non-cancer).
• Evaluate the model on the Kaggle Data Science Bowl 2017 dataset and compare with top competitors.

Proposed method

• Proposes a CNN with four convolution layers (11x11, 5x5, 3x3), two max-pooling layers, a fully connected layer, and a final two-unit softmax classifier.
• Uses ReLU activations after each convolution, followed by cross-entropy (softmax) loss and multinomial logistic regression objective.
• Trains with stochastic gradient descent (batch size 128, learning rate 0.001, momentum 0.9) on resized 120x120 CT slices.
• Evaluates performance using sensitivity, specificity, F1-score, and log-loss as used in Kaggle competitions.

Experimental results

Research questions

• RQ1 Can a compact deep CNN learn discriminative features from CT slices to distinguish lung cancer from non-cancer in a binary setting?
• RQ2 How does the proposed dCNN perform on the Kaggle Data Science Bowl 2017 dataset compared to leading competitors in terms of log-loss and standard metrics?

Key findings

• The proposed dCNN achieves sensitivity 0.87, specificity 0.991, and F1-score 0.95.
• The model attains a log-loss of 0.2098, outperforming top Kaggle competitors (e.g., grt123 at 0.3997).
• On the Kaggle dataset, the method shows strong discrimination with very high specificity, aided by a larger proportion of cancer-free images.
• Confusion matrix indicates effective cancer vs. non-cancer classification with favorable error distribution.