Among all cancer-related deaths, lung cancer leads all indicators, accounting for approximately 20% of all types. Patients diagnosed in the early stages have a 1-year survival rate of 81% to 85%, while in an advanced stage have 15% to 19% chances of survival. The primary manifestation of this cancer is through pulmonary nodule on computed tomography images. In the early stages, it is a complex task even for experienced specialists and presents some challenges to classify these nodules in benign or malignant. So, to assist specialists, computer-aided diagnosis systems have been used to improve the accuracy in the diagnosis. In this article, we explored and compared the use of random search, simulating annealing, and Tree-of-Parzen-estimators algorithms of hyperparameter tuning to find the best architecture of a convolutional neural network to classify small pulmonary nodules in benign or malignant with a diameter of 5 to 10 mm. Our best model used result was the model using the simulating annealing algorithm and yielded an area under the receiver operating characteristic curve of 0.95, the sensitivity of 82%, the specificity of 94%, and accuracy of 88% using a balanced data set of nodules. Therefore, our model is capable of classifying early lung nodules, where the patients have bigger chances of survival.

中文翻译：

---

利用卷积神经网络的超参数优化对小肺结节进行分类

在所有与癌症相关的死亡中，肺癌在所有指标中居首位，约占所有类型的 20%。早期确诊患者的 1 年生存率为 81% 至 85%，而晚期患者的生存机会为 15% 至 19%。这种癌症的主要表现是通过计算机断层扫描图像上的肺结节。在早期阶段，即使对于经验丰富的专家来说也是一项复杂的任务，并且对这些结节进行良性或恶性分类提出了一些挑战。因此，为了协助专家，计算机辅助诊断系统已被用于提高诊断的准确性。在这篇文章中，我们探索和比较了随机搜索的使用，模拟退火，和超参数调整的 Parzen 估计树算法，以找到卷积神经网络的最佳架构，以将直径为 5 至 10 毫米的良性或恶性小肺结节分类。我们使用的最佳模型结果是使用模拟退火算法的模型，使用平衡的结节数据集产生的接收者操作特征曲线下面积为 0.95，灵敏度为 82%，特异性为 94%，准确度为 88% . 因此，我们的模型能够对早期肺结节进行分类，其中患者的生存机会更大。使用平衡的结节数据集，灵敏度为 82%，特异性为 94%，准确度为 88%。因此，我们的模型能够对早期肺结节进行分类，其中患者的生存机会更大。使用平衡的结节数据集，灵敏度为 82%，特异性为 94%，准确度为 88%。因此，我们的模型能够对早期肺结节进行分类，其中患者的生存机会更大。