I. Introduction
According to the latest IARC report, in 2021, there are approximately 19.29 million new cancer cases worldwide, of which the United States accounted for about 11.8%, the second largest number of new cases of cancer worldwide [1]. Lung cancer is still far ahead of other types of cancer, accounting for about 11.4 percent of the total [2]. The report demonstrates that the incidence of lung cancer in rural areas will surpass that in cities, posing a threat to people's lives, mainly caused by active smoking, exposure to secondhand smoke, and radiation and harmful chemicals in the workplace [2], [3]. Overall, between 2001 and 2017, cancer deaths in the United States declined by an average of 1.5 percent per year. While the mortality rate has seen steady improvement the past few years, the low survival rate for lung cancer in 2019 of approximately 19% is attributed to its frequent late detection [4]. The methods that can provide patients for accurate pathological information based on lung pathological image to realize convenient and quick diagnosis are insufficient. CNNs is particularly good at automating feature extraction, with end-to-end learning rather than purely manual one. In the process of extracting pathological maps of lung tissue, there was an innovative Inception_v1 model concept that used 12 times fewer parameters than AlexNet to approximate the optimal density of the structure through dense components. The final result is three times faster than a similar network with a non-Inception architecture. Furthermore, HOG (Histogram of Oriented Gradients) is a manual method for extracting feature parameters. Its algorithm works by creating an image with gradient direction distribution, and then performing special normalization processing. This method can effectively detect the edge of lung cancer pathological images, even when the contrast is very low [5]. Therefore, it can help the doctor diagnose the condition more effectively. This paper outlines the combination of automatic feature extraction (inception_v3 network) and manual feature extraction (HOG and daisy) to classify lung cancer and normal tissue from pathological lung images.